Add optional mask & bias inputs with adaptive computation skipping

[Copilot]

This PR implements optional attn_mask and attn_bias inputs with adaptive computation skipping to improve performance and reduce unnecessary memory operations in Flash Dynamic Mask Attention.

Problem

The current implementation always assumes both attn_mask and attn_bias are active, causing:

• Unnecessary global memory loads when only one tensor is needed
• Needless dbias computation when no bias is conceptually required
• Inefficient workarounds like fake all-ones masks or zero bias tensors

Solution

Added support for 4 explicit modes with conditional processing:

Case	attn_mask	attn_bias	Behavior
A	None	None	Dense path, no block skip, no bias load/add (fastest)
B	Tensor	None	Block skip using mask, no bias add/dbias
C	None	Tensor	No block skip (all blocks active), add bias + compute dbias
D	Tensor	Tensor	Current behavior (mask skip + bias add + dbias)

Key Changes

Python Interface

• Both attn_mask and attn_bias parameters now accept Optional[Tensor] = None
• Added use_mask and use_bias flags passed to CUDA kernels
• Conditional gradient computation - dbias returned only when bias provided

CUDA Kernels

• Modified mask application logic to conditionally process mask and bias based on runtime flags
• Updated backward pass to skip dbias computation when use_bias=False
• All changes preserve performance when both tensors are provided

Usage Example

from flash_dmattn import flash_dmattn_func_auto

flash_attn = flash_dmattn_func_auto()

# Case A: Dense attention (fastest for dense workloads)
out = flash_attn(q, k, v, attn_mask=None, attn_bias=None)

# Case B: Sparse attention with mask only  
out = flash_attn(q, k, v, attn_mask=sparse_mask, attn_bias=None)

# Case C: Dense attention with bias (e.g., ALiBi, relative position)
out = flash_attn(q, k, v, attn_mask=None, attn_bias=position_bias)

# Case D: Full functionality (unchanged)
out = flash_attn(q, k, v, attn_mask=sparse_mask, attn_bias=position_bias)

Performance Benefits

• Case A: Eliminates both memory streams + skip logic overhead
• Case B: Removes bias path (saves reads/writes + math)
• Case C: Removes mask loads/OR reductions (simpler control flow)
• Case D: Baseline performance (unchanged)

Backward Compatibility

The implementation is fully backward compatible - existing code continues to work unchanged. Default parameter values maintain current behavior when not specified.

Fixes #161.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.