Fix GPU shader correctness and warnings surfaced by DXC warning flags

[num3ric]

• GpuShaderUtils: add intCast() helper that emits a C-style cast on HLSL (silences -Wfloat-conversion) and a functional cast on GLSL/MSL/OSL.
• FixedFunctionOpGPU: use intCast() for the three implicit float->int conversions in the ACES2 hue-weight setup and cusp-table search.
• FixedFunctionOpGPU: drop the unused _Add_SinCos_Shader calls in Gamut_Compress_Fwd/Inv_Shader (gamut-compress stays in JMh space, so cos_hr/sin_hr were dead).
• GradingBSplineCurve: in the HUE-FX inverse path (curveIdx == 7), offset knStartY by knStart to match the CPU evalCurveRevHue.
• GradingRGBCurveOpGPU: inverse eval no longer declares identity_x (it's never used in the inverse), and the in/out qualifier branch is folded into a single emitter — matches the CPU evalCurveRev signature.
• DxApp: when verbose (test_gpu_exec --dx -v), pass -Wall -Wextra -Wconversion to DXC and print DXC_OUT_ERRORS even on success, so generated-shader warnings are visible during debugging.

[num3ric]

See attached file for warning logs (without fixes included here): dxc_warnings_full.log

[fallenmi]

Reviewed exact head 1d41e95d8738fdccc3d4bac4104c665c5bcbc519 against exact PR base 57a94a8003c00be506a41f54a080d9fefaba6d22 and current main 5a808fb57a94c7229640a97835c420c9a1fbd1fe. Although GitHub marks the branch behind, it merges cleanly; I tested the resulting current-main tree a28f55e4e14a070f2a82f779bb73c0de038a9be9.

The shader changes are internally consistent. The HLSL-only C-style integer cast preserves existing truncation while avoiding DXC's functional-cast warning; other shader languages retain their required functional syntax. The HUE-FX inverse lower bound now matches the CPU path's knot offset, the inverse RGB-curve helper safely drops an argument that was redundant because its empty-curve path returns x, and the removed sine/cosine setup was dead in the standalone JMh gamut-compress paths.

Exact base, head, and current merge all build successfully. The same 50 shader/curve/fixed-function focused tests pass on all three. The full CPU runs are also baseline-equivalent: base and head each have 1144 passes / 4 failures, and current merge has 1149 passes / the same 4 failures (two unavailable-monitor failures and two pre-existing AppleClang floating-point tolerance failures). Live exact-head CI is fully green: 21 check-runs plus EasyCLA and ReadTheDocs, with no pending, failed, or action_required workflow. I found no blocking issue.

AI disclosure: I used OpenAI Codex to inspect the exact revisions and repository policy, build the exact base/head/current-main merge oracle, run and compare the focused and full CPU suites, verify live checks and review history, and draft this review. I verified the evidence and conclusion.