head, seq, batch grid order for triton flash attention bwd.

[Chi-Chu319]

Change the grid order for triton flash attention bwd.

The grid is changed to head, seq, batch. It provides the following benefits

1. having head as the first dim with the assumption of head % 8 (number of XCD) == 0 keeps each entire sequence on one XCD so good causal load balancing between XCDs. and good sequence locality
2. having the batch as the last makes sure all the XCD gets all the batches so the THD load balancing is achieved.

suppose we have head // num_XCD = x
each XCD would get (xcd_id * x + (xcd_id + 1) * x, ..., ...), with the grid (head, seq, batch) (we know that the first dim moves the first).

perf on Mi300:
main_perf:

this branch:

perf on mi350:
main_perf:

this branch: