Feature Request: Add THD / Packed-Sequence Training Support

Nobody has claimed this yet.

### 🚀 The feature, motivation and pitch

## Motivation

Some training frameworks represent variable-length batches in a packed THD layout, where activations are shaped like `(total_tokens, num_heads, head_dim)`, or more generally `(total_tokens, hidden_dim)`, instead of the standard padded batch layout `(batch_size, seq_len, ...)`.

This layout is useful for sequence packing and variable-length training because it avoids padding overhead.

Supporting this layout may require passing sequence metadata such as `cu_seqlens` or `max_seqlen` to kernels that need per-sequence position information.

## Request

Add support for THD / packed-sequence training layouts across Liger kernels.

The long-term goal is to support inputs such as `hidden_states: (total_tokens, hidden_dim)` and attention tensors such as `(total_tokens, num_heads, head_dim)`, with metadata such as `cu_seqlens` and `max_seqlen` where needed.

### Alternatives

_No response_

### Additional context

_No response_

The issue names no implementation files, entry points, or tests. Start by surveying the existing Liger kernels and their shape-related tests, then determine which kernels need THD inputs and sequence metadata such as cu_seqlens and max_seqlen. Done means the supported kernels accept packed layouts without padding assumptions and have coverage for variable-length training.

Written by the indexing model from the issue text.