perf: move to seq-parallel atomic reduction in bwd_b

[TimoImhof]

The bwd_b kernel is currently the primary bottleneck in the fused backward pas:

It iterates over the sequence dimension ($S$) within a single thread block:

# dL/d(out).T @ (X @ A.T)
    for s in range(0, tl.cdiv(dL_dout_S, BLOCK_S)):
        # ... inner = X @ A.T ...
        # ... _total_acc = tl.dot(dL_out, _interm_tile) ...

leading to redundant memory traffic: Both $X$ and $A$ are re-loaded from global memory $S/BLOCK_S$ times. Since $S$ is typically the largest dimension, this creates massive, unnecessary VRAM pressure.

Proposed Fix: Sequence-Parallelism Mirroring the successful optimization of bwd_a - this PR refactors bwd_b to:

• Parallelize over $S$: Map the grid to the sequence dimension so that multiple blocks process chunks of the sequence concurrently.
• Eliminate Redundant Loads: Each block loads its chunk of $X$ and $A$ exactly once.
• Atomic Accumulation: Uses tl.atomic_add to aggregate partial gradients into the final $B$ matrix.