Test comparing attention weights fails for FlashInfer if `q_len=1`

[mseeger]

Describe the bug

Look at test/attention/test_attn_weights.py, test_larger_comparison, the q_len=1 cases.

Currently, FlashInfer is skipped for this one. If the skip is removed (in get_variants), the tests fail for FlashInfer. All other variants give the same results.

FlashInfer works for all other q_len values. Likely this is because the case q_len=1 is implemented differently.

[mseeger]

Note: Please wait for #95 to be merged!

[mseeger]

OK, this IS a problem: For now, our evaluation scripts refuse to use FlashInfer SDPA, because the q_len=1 case is needed for token generation, and generation fails right now (giving very poor results).

Training is not affected, since chunks are never just size 1.