Add standalone usage example and inference-only note

Problem

The README only shows integration via FLA's chunk_kda interface. There is no standalone example showing direct usage of the flash_kda.fwd API, and no minimal reproduction script.

Recommendation

Add examples/basic_usage.py:

import torch
from flash_kda import fwd

B, H, T, D = 2, 8, 1024, 128
q = torch.randn(B, H, T, D, dtype=torch.bfloat16, device='cuda')
k = torch.randn_like(q)
v = torch.randn_like(q)
g = torch.randn(B, H, T, dtype=torch.float32, device='cuda')
beta = torch.rand(B, H, T, dtype=torch.float32, device='cuda').sigmoid()
scale = D ** -0.5

out = fwd(q, k, v, g, beta, scale)
print(f"Output shape: {out.shape}")  # [2, 8, 1024, 128]

Also add a note in README that this is inference-only (no backward pass) — users needing training should fall back to FLA's Triton path.

Impact

Documentation. Reduces friction for new users who want to try the kernel directly without understanding FLA's internals.

Start by reading the README and the existing FLA integration to match project conventions. Add examples/basic_usage.py using the documented flash_kda.fwd API, then add the inference-only note to the README, including the fallback to FLA's Triton path for training. Done means the standalone example is present and the README clearly explains the limitation.