Enable softmax-sink FMHA kernels for TransformerEngine

[shurale-nkn]

Motivation

Required by ROCm/TransformerEngine#678 — [ROCm] Jax Add softmax sink (learnable
off-by-one) support for the ROCm/CK fused attention backend.

Receipt 700's forward rule carried cond &= kernel_ctx.pipeline.F_sink == "f", so
codegen emitted zero sink instances and a forward call with has_sink=true found no
matching trait — CK's generated fmha_fwd() returned -1, surfacing in TE as
fused attn configs not supported in ck_fused_attn fwd pass. That invariant was written
when TE had no sink support, so it is now stale. This PR drops the F_sink condition
from 0003-ck-te-receipt-700.patch and updates the two spots that document the
invariant (patch preamble and the generate.py help text).

Also adds 0008-fmha-bwd-sink-fix.patch: CK's fmha_bwd_kernel.hpp indexed sink_ptr
as i_batch * nhead + i_nhead, but the buffer holds one value per head (shape
[nhead]), matching how the forward kernels already read it, so i_batch must not
factor into the offset.

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.