Add 2048-token pages to FMHA batch prefill

[rlrs]

Proposed changes

Add 2048-token pages to the FMHA batch-prefill generator.

vLLM uses 2048-token cache pages for hybrid models whose recurrent layers have
large state-alignment requirements. The batch-prefill kernel already accepts a
compile-time page size, so this only adds 2048 to the generated set.

On an MI250X with a 128K context, the generated kernel was 1.70x faster than
the existing Triton paged-prefill path. The same comparison at 256K was 1.69x.

This does not change kernels generated for existing page sizes or for non-gfx9
targets.

Validation:

python example/ck_tile/01_fmha/generate.py \
  -d batch_prefill -r 200 -f '*ps2048*' -o <output>

1681 generated files, all containing the 2048 page-size specialization

I also checked the generated kernel against the PyTorch attention reference for
the vLLM BF16 MQA shape and ran an end-to-end 128K-context serving benchmark.

Checklist

• I have added tests relevant to the introduced functionality, and the unit tests are passing locally
• I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, IF the test takes more than 30 seconds to run.
• I have added inline documentation which enables the maintainers with understanding the motivation
• I have removed the stale documentation which is no longer relevant after this pull request
• (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
• I have run clang-format on all changed files
• Any dependent changes have been merged

Discussion

This change was prepared with AI assistance and reviewed and tested by the
submitter.