fix: expose alpha param through LigerFusedLinearDPO public API

[micdoh]

Summary

• `LigerFusedLinearDPOFunction.forward()` was missing the `alpha` parameter entirely, so the NLL loss scaling weight silently defaulted to `1.0` no matter what callers passed — the parameter exists in `LigerFusedLinearPreferenceBase` but was never plumbed through `LigerFusedLinearDPOFunction`.
• `LigerFusedLinearDPOLoss` (the `nn.Module` wrapper) similarly had no `alpha` argument.
• Added one extra `None` to `backward()` to match the new positional count.
• Fixed positional-arg ordering in the two existing functional tests that called `LigerFusedLinearDPOFunction.apply(...)` directly.
• Added `test_alpha_scales_nll_loss` to verify `alpha` actually reaches the loss computation.

Changes

`src/liger_kernel/chunked_loss/dpo_loss.py`

• Add `alpha: float = 1.0` to `LigerFusedLinearDPOFunction.forward()` and forward it to `super().forward()`
• Add one `None` to `backward()` return (was 11, now 12 — one per non-tensor arg)
• Add `alpha: float = 1.0` to `LigerFusedLinearDPOLoss.init()` and `self.alpha` assignment
• Pass `self.alpha` to `LigerFusedLinearDPOFunction.apply()` in `LigerFusedLinearDPOLoss.forward()`

`test/chunked_loss/test_dpo_loss.py`

• Insert `1.0` (alpha) in the correct positional slot in `test_correctness_functional` and `test_correctness_functional_apo_loss_types`
• Add `test_alpha_scales_nll_loss` regression test

Test plan

• Existing `test_correctness`, `test_correctness_functional`, `test_correctness_apo_loss_types`, `test_correctness_functional_apo_loss_types` all pass
• New `test_alpha_scales_nll_loss` passes and confirms `alpha != 1.0` changes the loss value