graphics: rewrite screenshot and api capture image processing (#870)

## Link to GitHub Issue or related Pull Request, if one exists
#0

## Description of change
Significantly speeds up API screen capture and D3D9 screenshots saving.
Two reasons for doing this:

1. We now have a 4K game (GITADORA) and existing capture code was taking
multiple seconds.
2. Renewed user interest on streaming as we have a couple more companion
apps in active development.

**API screen capture (streaming), 1280x720:** 14.3ms -> 6.3ms per frame.
Back buffer copies go to pooled `D3DPOOL_SYSTEMMEM` surfaces via
`GetRenderTargetData` instead of allocating a lockable render target
every frame, and TooJpeg is replaced with libjpeg-turbo (encode 9.8ms ->
3.0ms). MSAA remains unsupported

**Screenshots for GITADORA arena model, across 4 screens with one of
them 4K**: 4068ms -> 124ms. `D3DXSaveSurfaceToFileA` is replaced with
fpng (encode 4043ms -> 76ms) and the screens encode in parallel.
Dropping D3DX also removes the `d3dx9_43.dll` ... `d3dx9_24.dll` probing
loop, so screenshots no longer fail outright on machines with no D3DX9
runtime installed.

Screenshot surfaces are read on the present thread, so no D3D call
reaches another thread for screenshots. This fixes a hang in DDR X2
introduced earlier in the branch: its device has no internal locking,
and reading the surface on a pool thread while the present thread sat
inside `GetRenderTargetData` left the game's own render thread
deadlocked.


## Testing

- **GITADORA** (arena model, D3D9Ex, 4K main plus three subscreens,
windowed) with
`-screenshotsub`: three sets of four screenshots, images verified
correct. Completion
order differs between sets, so the screens really are encoding in
parallel.
- **LovePlus** (KLP, plain D3D9, 768x1360): covers the inline path used
by games whose
  image processing must not leave the present thread. 
- **API screen capture** through a companion app: live video correct
throughout.
- **Print Screen** bound as the screenshot key: the clipboard copy
succeeded on every shot.
- Quitting the game after capturing leaves no `IDirect3DDevice9`
reference count warning,
  so the pooled readback surfaces are released along with the device.

src/spice2x/external/libjpeg-turbo/jdcoefct.h

@@ -0,0 +1,88 @@
+/*
+ * jdcoefct.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2020, Google, Inc.
+ * Copyright (C) 2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+#define JPEG_INTERNALS
+#include "jpeglib.h"
+
+
+#if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
+
+/* Block smoothing is only applicable for progressive JPEG, so: */
+#ifndef D_PROGRESSIVE_SUPPORTED
+#undef BLOCK_SMOOTHING_SUPPORTED
+#endif
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_coef_controller pub; /* public fields */
+
+  /* These variables keep track of the current location of the input side. */
+  /* cinfo->input_iMCU_row is also used for this. */
+  JDIMENSION MCU_ctr;           /* counts MCUs processed in current row */
+  int MCU_vert_offset;          /* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;    /* number of such rows needed */
+
+  /* The output side's location is represented by cinfo->output_iMCU_row. */
+
+  /* In single-pass modes, it's sufficient to buffer just one MCU.
+   * We allocate a workspace of D_MAX_BLOCKS_IN_MCU coefficient blocks,
+   * and let the entropy decoder write into that workspace each time.
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays; it is used only by the input side.
+   */
+  JBLOCKROW MCU_buffer[D_MAX_BLOCKS_IN_MCU];
+
+  /* Temporary workspace for one MCU */
+  JCOEF *workspace;
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+#endif
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  /* When doing block smoothing, we latch coefficient Al values here */
+  int *coef_bits_latch;
+#define SAVED_COEFS  10         /* we save coef_bits[0..9] */
+#endif
+} my_coef_controller;
+
+typedef my_coef_controller *my_coef_ptr;
+
+
+LOCAL(void)
+start_iMCU_row(j_decompress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row (input side) */
+{
+  my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (cinfo->input_iMCU_row < (cinfo->total_iMCU_rows - 1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->MCU_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+#endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */