fix(graphics): apply YCbCr->RGB fixed-point descale in progressive decoder/encoder

[Rocco De Angelis (rdeangel)]

Problem

Against some real-world EGFX servers (xrdp, GNOME Remote Desktop), the progressive (RemoteFX) decoder produces posterised output. Smooth colour gradients collapse to a handful of saturated primaries and luma loses its mid-tones. Spatial detail (text, icons) stays crisp, which points away from a scaling issue and toward colour conversion.

Root cause

reconstruct_to_rgba() passes the dequantized Y/Cb/Cr values into the BT.601 YCbCr->RGB conversion, but the fixed-point headroom that dequantize_component_ccq intentionally retains isn't removed before conversion.

dequantize_component_ccq applies the spec-defined << (quant_value - 1) shift. For the wire quant table this works out to +5 bits of headroom, which matches FreeRDP's reference decoder (see the /* -6 + 5 = -1 */ comment in libfreerdp/primitives/prim_colors.c). The colour-conversion stage is the natural place to remove that headroom (R = ((CrR + Y) >> divisor) >> 5). This PR adds the missing >> 5 and the corresponding 128 << 5 = 4096 luma offset. Without them, each channel exceeds the legal ±128 range and clamp_u8 pins it to 0 or 255.

Worth noting: this isn't the integer inverse DWT. The DWT is unity-gain, verified separately below.

Verification

1. DWT unity-gain check. A flat 64×64 tile of value 100 forward-transforms with no coefficient exceeding ~100 and round-trips back to exactly 100, so the ×2^5 scale doesn't originate there. Added tests/dwt_gain.rs to lock this in.
2. Live server quant trace. Against an xrdp stream, the per-band << (q-1) factors land at +5 bits of headroom (max coefficient scaled ~32×). The +5 bits come from the dequant table.
3. FreeRDP cross-check. libfreerdp/codec/progressive.c and libfreerdp/primitives/prim_colors.c apply the same << (q-1) shift and the matching >> 5 removal at the colour stage, so this patch's math lines up with the reference.
4. Visual confirmation. With the >> 5 descale applied, xrdp and GNOME RD streams render with correct colour gradients; without it, the same streams are posterised.

Scope of this PR

Kept symmetric so the crate's internal encode<->decode round-trip test stays identity:

• Decoder: >> 5 plus +128 luma offset in reconstruct_to_rgba (matching FreeRDP).
• Encoder: matching << 5 plus offset on the RGB->YCbCr path in rgba_to_ycbcr, so the round-trip doesn't regress.

Files

• crates/ironrdp-graphics/src/progressive.rs — decoder descale and encoder scale-up
• crates/ironrdp-graphics/tests/dwt_gain.rs — regression test confirming the DWT is unity-gain