Add NVFP4 per-token quantization recipe

[cael-ling]

Description

This PR adds an NVFP4 per-token quantization fast path for bf16 inputs on Blackwell (SM100+) for Model Pre-training. Per-token uses per-row / per-column outer amax instead of the per-tensor scalar amax, which factors cleanly out of the GEMM K-summation and lets the inner GEMM stay on production cuBLASLT NVFP4 plus a thin trailing post-scale -- no CUTLASS fork required.

Status: draft. The cast kernel and GEMM composite, byte-equal-verified against a Python reference, and benched against the per-tensor (RHT + SR) recipe are still in progress. Partial experimental results are shown as follows.

Cast-only (fwd/bwd: amax+cast) performance comparisons in Eager mode:

M	K	N	per-token (us)	per-tensor (us)	per-token/per-tensor
128	4096	4096	61.4	106.8	0.575
512	4096	4096	58.7	101.4	0.578
1024	4096	4096	57.1	105.8	0.540
4096	4096	4096	62.3	101.9	0.612
8192	4096	4096	61.6	104.9	0.587
16384	4096	4096	86.0	118.2	0.728
32768	4096	4096	146.5	198.0	0.740

Cast-only (fwd/bwd: amax+cast) performance comparisons in Graph mode:

M	K	N	per-token (us)	per-tensor (us)	per-token/per-tensor
128	4096	4096	16.7	28.9	0.579
512	4096	4096	17.8	28.0	0.637
1024	4096	4096	19.5	37.0	0.527
4096	4096	4096	28.9	39.2	0.737
8192	4096	4096	46.5	67.3	0.691
16384	4096	4096	80.3	120.2	0.668
32768	4096	4096	138.7	199.8	0.694

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes