[FlexAttention] add support for learnable biases in Inductor (pytorch…

…#137452)

# Summary

The follow up PR to: pytorch#137526.  In this pr, we actually update the lowerings for the flex_attention backwards kernel to generate fused backward gradient calculations for any captured buffers that require grads.

We are doing this using tl.atomic_add to scatter the correct gradients into zeroed out buffer for any captured buffers that required grads. Added many test cases and found.  Along the way found some masking bugs.

There are likely some performance cliffs here, specifically with D-types and on different GPUs. Planned to do this in a follow-up and profile the current strategy. We are explicitly choosing reduced memory over increased performance right now.

By using atomics, we do not need to realize a full attention scores matrix. However, this comes with two downsides. One, this is potentially slower in some cases, and two, the gradient calculation for any captured buffers is non-deterministic.

## Worked Example

Lets do the case where you are reading from one bias that doesn't require grad and using this to index into another that does.

ScoreMod:
```Python
bias = torch.randn(
    params.seq_length,
    device=self.device,
    dtype=params.dtype,
    requires_grad=True,
)

offset = torch.randint(
    0,
    params.seq_length,
    (params.seq_length,),
    device=self.device,
)

def score_mod(score, b, h, q_idx, kv_idx):
    return score + bias[offset[q_idx]]

```

I am removing all but the new subgraph injected into the backwards:

``` Python
    dsT = pT * (dpT - Di[None, :])
    # ~~~~~~~~~~~~~~~~~~~ Apply joint modification  ~~~~~~~~~~~~~~~~~~~
    grad_scores = (dsT)

    # ~~~~~~~~~~~~~~~~~~~ Apply other buffer grad writes ~~~~~~~~~~~~~
    idx_b = off_z
    idx_h = off_hq
    idx_m = m
    idx_n = n
    scatter_mask = offs_m1[None, :] < Q_LEN and offs_n1[:, None] < KV_LEN
    tmp4 = (dsT).to(tl.float32)
    tl.atomic_add(out_ptr1 + (tl.broadcast_to(tl.load(in_ptr16 + idx_m), tmp4.shape)), tmp4, scatter_mask, sem='relaxed')

    # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
```
## Key points
* We always accumulate to float 32 grad buffers regardless of the type in the forward. This is because we normally do all computation intra kernel w/ fp32 accumulation and we want the same behavior for atomic additions
* We are currently restricted to 1 scatter in the kenrel. I have some ideas on fx rewrites that would remove this restrictions but for now have nice error message w/ work around and will leave as a follow up.
* Will do more extensive performance/ memory profiling in a follow up.

### Toy E2E example
I have a toy E2E training example PR in the gym for now: pytorch-labs/attention-gym#84
I plan to update to a realistic learnable bias before landing

Pull Request resolved: pytorch#137452
Approved by: https://github.com/Chillee