fix(kernel): correct fp4 cache scale rounding

[LZ-QWQ]

Based on the analysis and proposed fix in #673, fix DSV4.1 FP4 cache scale rounding reported on RTX 3090 Ti and reproduced on Radeon GPUs and RTX4090 with software emulation.

The compiler can fuse reciprocal multiplication with the emulated e4m3 magic-add into an FMA, skipping intermediate FP32 rounding and changing subnormal ties. Use tl.math.div_rn(amax, 6.0) only for software emulation; retain the existing multiplication for native conversion through compile-time e4m3_native_cx().

Validation:

python -m pytest tests/kernels/test_dsv41_pack.py

GPU / path	Before	After
W7900D	8 passed, 2 failed	10 passed
R9700	8 passed, 2 failed	10 passed
RTX4090 default	10 passed	10 passed
RTX4090 emulated	8 passed, 2 failed	10 passed

For RTX4090 emulation, prefix the command with FREETOKEN_FORCE_E4M3_EMU=1. KV pool and sparse-attention tests also pass (15/15 per path).

AMD validation used ROCm 10.1 with #535's torchvision prerelease lower bound.

The checked RTX4090 default specializations produce byte-identical executable code before and after the change.

[jason-fxz]

LGTM.