perf(cuda): add calibrated mixed precision and fuse diffusion linears

[elprofesoriqo]

Summary

• add an explicit approximate FP16-deep diffusion mode while keeping exact FP32 as the default
• calibrate the mixed-precision layer set against matched PushT seeds and reject the quality-regressing seven-layer candidate
• fuse 1xK linear bias and optional Mish on CUDA, reducing the ten-step launch census from 1,000 to 850
• expose precision selection through the C ABI, Python extension, LeRobot adapter/config, rollout CLI, and evaluators
• check in complete alternating timing/parity samples and PushT calibration results

Evidence

GTX 1650 4 GB, WSL2, Release CUDA, pinned diffusion_pusht checkpoint:

• approximate mode gate: 104.216 ms to 82.620 ms median native p50 before fusion, 20.72% lower, all 5 paired trials won
• matched PushT seeds 7-16: exact 8/10 success, approximate 8/10; mean max reward 0.998138 vs 0.998099
• rejected calibration: converting the seventh deep convolution reduced success to 7/10
• exact fusion gate: 105.815 ms to 104.257 ms median p50, 1.47% lower, all 5 paired trials won, bitwise-identical actions
• post-fusion exact vs approximate: 98.736 ms vs 83.574 ms median native p50

This is PushT simulation evidence. SO-101 task success is not yet measured. The 10 ms target is not met.

Verification

• scripts/verify_local.ps1 passed on the final tree
• Windows native CTest: 85/85
• focused native kernels: 19/19
• LeRobot unittest suite: 110 passed, 1 expected artifact-evidence skip in the canonical run
• CUDA focused C ABI and Diffusion tests: 16/16
• CUDA fused linear reference test passed
• Release CUDA and WSL CPU builds passed
• diffusion parity remained within the existing exact-mode gates

Dependencies

#202 and #203 are merged. This PR is rebased onto current main and contains one new commit.