[Cute,Sm100,Bwd] refine bwd swizzle for deterministic

[jayhshah]

This PR fixes the varlen swizzle num_n_blocks from num_m_blocks calculation for 2cta and includes dqaccum in the head size for not spilling l2. This noticeably improves backward deterministic FLOPs (for non-varlen, limited to non-causal but doesn't appear to cause regression for causal; for varlen, improve perf across the board thanks to the fix).

Example benchmarks:

non-varlen deterministic, MHA nheads = 16

hdim causal batch seqlen     PR          		MAIN
----------------------------------------------------------------
128  False     4   8192      4.58/1200/53.3% 	4.75/1157/51.4%
128  False     2  16384      9.32/1180/52.4%	10.08/1091/48.5%
128  False     1  32768     19.69/1117/49.6%	20.48/1074/47.7%

128   True     4   8192      2.50/1100/48.9%	2.50/1100/48.9%
128   True     2  16384      4.60/1195/53.1%	4.60/1195/53.1%
128   True     1  32768      9.74/1129/50.2%	9.86/1115/49.6%

varlen deterministic, MHA nheads = 16

hdim causal batch seqlen     PR          		MAIN
----------------------------------------------------------------
128  False     4   8192      4.66/1180/52.5%	5.13/1071/47.6%
128  False     2  16384      8.97/1225/54.5%	10.15/1083/48.1%
128  False     1  32768     19.45/1131/50.3%	20.85/1054/46.9%

128   True     4   8192      2.59/1061/47.1%	2.83/973/43.2%
128   True     2  16384      4.72/1166/51.8%	5.06/1086/48.3%
128   True     1  32768      9.72/1131/50.3%	10.09/1089/48.4%

[tridao]

does this affect any of the 2cta bwd code?

[jayhshah]

This change is meant to get the right head swizzle heuristic for 2cta bwd, by accounting for the num_m_blocks being defined with respect to tiler divided by the cluster shape.

num_n_blocks here is only used to derive nheads_in_l2, so it doesn't affect correctness.

[tridao]

For the tile_shape_mn[0] do we pass the CTA's tile shape or the cluster tile shape? I think we had this discussion and realized we have not been consistent.
In any case if it doesn't affect correctness it's fine w me.

[jayhshah]

Currently for the tile scheduler args we pass cta_tiler[:2] as tile_shape_min and cluster_shape_mn as a separate parameter; this makes the most sense to me since cluster shape is in principle separate from use of 2cta mma.