feat: add SM120 fmha_v2 kernels to AOT pip wheel builds (#2885)

## Summary

`gen_trtllm_fmha_v2_sm120_module()` exists in `jit/attention/modules.py`
and the JIT runtime path (`generate_kernels.py`) already dispatches to
it correctly. However, `aot.py`'s `gen_all_modules()` — which drives the
pip wheel AOT build — was missing it from the `has_sm120 or has_sm121`
section.

This means SM120/SM121 devices using a pip wheel would never get the
fmha_v2 SM120 kernels compiled into the wheel, and would have to fall
back to slower paths.

**Fix:** Add `gen_trtllm_fmha_v2_sm120_module()` to the `has_sm120 or
has_sm121` block in `aot.py`, alongside the other SM120 modules (fused
MOE, GEMM, FP4 quantization).

No behavior change for JIT users; only affects AOT pip wheel builds.

Addresses the AOT gap noted in #2555.

Contributed by Second Nature Computing (https://joinsecondnature.com)

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Chores**
* Expanded optimized inference module support for SM120 and SM121 GPUs
to include attention kernels in addition to existing fused MoE and GEMM
optimizations.
* Increased runtime coverage and readiness for attention-heavy workloads
on those architectures, improving performance consistency for models
using attention.

[![Review Change
Stack](https://storage.googleapis.com/coderabbit_public_assets/review-stack-in-coderabbit-ui.svg)](https://app.coderabbit.ai/change-stack/flashinfer-ai/flashinfer/pull/2885)
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>