fix(kernels): remove global --use_fast_math; selective per-kernel intrinsics only (T3.1)

[dndungu]

Implements zerfoo plan-gpu-training-hardening T3.1 + S3.1.1 (verifies UC-GH-5).

What

• Drop --use_fast_math from NVCC_FLAGS (internal/cuda/kernels/Makefile). The flag swapped every expf/tanhf/logf/division for unsafe approximations library-wide; the non-saturating fast tanhf caused the GELU blowup behind the GPU f32 CrossAsset cliff (fix(cuda): saturate kernel_tanh argument (--use_fast_math tanhf overflow) #125).
• Selective intrinsic, oracle-gated: kernel_softmax keeps __expf after max-subtraction (argument <= 0, result in (0,1], ~2^-21 rel error — inside the Softmax tolerance atol 1e-6 / rtol 1e-4). Everything else uses accurate libdevice functions.
• Tanh saturation clamp (fix(cuda): saturate kernel_tanh argument (--use_fast_math tanhf overflow) #125) RETAINED as defense-in-depth.
• oracle-gen -engine gpu: GenerateAllWith records compute.GPUEngine bundles through the same format — the GPU-side oracle gate the scripts/oracle README reserved for T3.x.
• compute/gpu_kernel_bench_test.go: per-kernel GPU micro-benchmarks (Exp/Log/Sqrt/Rsqrt/Sin/Cos/Tanh/Div/Softmax) for the before/after perf delta.

Gates (GB10 via Spark, serialized — results to be appended to this PR + devlog)

• sm_121 libkernels.so rebuilt from this branch
• GPU-engine oracle bundles vs torch (run_oracle.py, NGC 26.02): all ops green within tolerance
• Parity suite (scripts/parity/run.sh, test(parity): Wolf-pattern training loop under poison -- GB10 + CI variants (S2.3.1) #140 pattern): both schedules + red-proof green against the rebuilt kernels
• Per-kernel + end-to-end perf delta recorded in devlog (a regression is expected and acceptable — correctness is the point)

[dndungu]

GB10 gate results (all green)

Build: sm_121 libkernels.so rebuilt from 393d8eb (Spark pod ztensor-t31-build-393d8eb, NGC nvcc).

Oracle (S3.1.1) — first live GPU-engine oracle run: 25 GPU-engine bundles (oracle-gen -engine gpu) vs torch 2.11.0a0+nv26.02 on the GB10: 25 passed / 0 failed / 0 errored. Softmax with the selective __expf: fwd max_abs 6.0e-8 / max_rel 1.4e-7 (tolerance 1e-6 / 1e-4). Tanh exact (0.0/0.0). Report archived at /home/ndungu/t31/393d8eb/report.json.

Parity sanity (#140 pattern, rebuilt kernels): TestParity_GPUvsCPU_ArenaStressSchedules_GPU PASS (both schedules), TestParity_GPURedProof_GPU PASS, TestTrainingLoop_WolfPattern_GPU PASS. Baseline run with the OLD fast-math kernels also green (constant gate).

Perf delta (engine-level, 4M f32, mean of 3x100):

kernel	fast-math µs	no-fast-math µs	delta
Exp	425.6	417.1	-2.0%
Log	419.7	423.9	+1.0%
Sqrt	417.9	420.3	+0.6%
Rsqrt	417.9	426.1	+2.0%
Sin	419.3	423.3	+0.9%
Cos	421.7	427.9	+1.5%
Tanh	417.7	421.1	+0.8%
Div	778.4	778.9	+0.1%
Softmax	428.2	430.6	+0.6%

End-to-end: parity suite 0.720s (old) → 0.691s (new). All deltas within run-to-run noise (~±2%): the engine path is bandwidth-bound at these sizes, so no measurable regression. Full numbers + ops notes in the devlog entry in this PR.

[dndungu]

Post-merge-of-main re-gate (pod ztensor-t31-gate-new2-dbcb0da, merged head dbcb0da, rebuilt no-fast-math kernels): the FULL current parity gate including the #141 attention stress test is green on the GB10 — TestAttentionTraining_WolfPattern_GPU PASS, TestParity_GPUvsCPU_ArenaStressSchedules_GPU PASS, TestParity_GPURedProof_GPU PASS, TestTrainingLoop_WolfPattern_GPU PASS (VALIDATION_OK_new2).

[dndungu]

Superseded by #143 (same change + GB10 evidence), which is merged — main has 1fd2e89 + b596563.