Is the combination of var-len, paged KV and split KV supported?

[masahi]

Support for var-len + paged KV was added in #831 and I confirmed that it works when num_splits = 1. But if I force the kernel to use num_splits > 1, I get incorrect results.

Looking at the code, I suspect var-len + paged KV + split KV is not currently supported. First, the storage representation of partial reductions is very wasteful for the var-len case:
https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/flash_api.cpp#L316

Second, the way the output pointer is indexed in combine_attn_seqk_parallel seems to assume the fixed query-length format (batch_size x seqlen_q x num_heads x head_size). Shouldn't we use binfo.q_offset there?
https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash_fwd_kernel.h#L1273-L1274

@tridao @sgrigory