Recompute attention-residual

[ngc92]

This requires one more matrix multiplication on the backward pass, but halves the amount of activation memory that needs to be kept compared to the current setting.

This allows us to achieve reasonable mfu even with the 14B model on 4x4090:

Old:

Model	nGPU	DType	Batch	TPS	SOL	TTB
Qwen2.5-14B⁶	4	fp8	4	3.2k	22%	87h
Qwen2.5-14B⁷	4	bf16	4	2.5k	33%	111h

New:

Model	nGPU	DType	Batch	TPS	SOL	TTB
Qwen2.5-14B⁸	4	fp8	8	6.0k	42%	47h
Qwen2.5-14B⁹	4	bf16	8	4.5k	58%	62h