Add Pythonic API wrapper with automatic workspace management

Problem

The C++ function requires an externally-managed workspace tensor, but the Python flash_kda/__init__.py is essentially empty — it just imports the C extension with no Pythonic wrapper.

Users currently need to:

size = flash_kda.get_workspace_size(B, H, T, CHUNK)
workspace = torch.empty(size, dtype=torch.uint8, device='cuda')
flash_kda.fwd(q, k, v, g, beta, scale, out, workspace=workspace, ...)

Recommendation

Add a Pythonic wrapper that handles workspace allocation automatically:

def fwd(q, k, v, g, beta, scale, out=None, **kwargs):
    """Forward pass for Kimi Delta Attention.
    
    Args:
        q: [B, H, T, D] query tensor (bf16/fp16)
        k: [B, H, T, D] key tensor
        v: [B, H, T, D] value tensor
        g: [B, H, T] gate logits (fp32)
        beta: [B, H, T] beta modulation (fp32)
        scale: float, typically 1/sqrt(D)
    
    Returns:
        out: [B, H, T, D] output tensor
    """
    if out is None:
        out = torch.empty_like(q)
    size = get_workspace_size(q.shape[0], q.shape[1], q.shape[2], 16)
    workspace = torch.empty(size, dtype=torch.uint8, device=q.device)
    _fwd_impl(q, k, v, g, beta, scale, out, workspace, **kwargs)
    return out

Also add docstrings with expected shapes and dtypes, and define __all__ for clean from flash_kda import * behavior.

Impact

Usability. Makes the API accessible to users who don't want to manage CUDA workspace allocation manually. Reduces boilerplate and potential for OOM errors from incorrect size calculations.

Start in flash_kda/init.py and inspect the existing C extension entry points, including get_workspace_size and fwd. Add the documented Python wrapper, optional output allocation, automatic workspace allocation, and all; done means callers no longer need to create workspace tensors manually while the expected shapes and dtypes are documented.