Sequence Packing Bug: cu_seqlens not propagated to AutoModel

**Describe the bug**

When sequence packing is enabled with AutoModel custom/registered models (e.g., GLM-4.7-Flash, Qwen3.5-MoE), cu_seqlens is silently dropped due to an interface mismatch between NeMo-RL and AutoModel. This causes cross-sequence attention leakage in packed batches — the model attends across sequence boundaries as if the entire packed tensor were a single sequence.

**Steps/Code to reproduce bug**

Trigger two runs with an AutoModel registered model, one with sequence packing enabled, one not.
Compare the `gen_kl_error`. The symptom is the value of `gen_kl_error` gets much higher when packing is enabled. e.g. On glm4.7-flash, from 0.x down to 0.00x.

All models registered in AutoModel's `ModelRegistry` are affected. As sequence packing is enabled by default. It is of high impact.

**Additional context**

Root cause:

NeMo-RL constructs a `FlashAttentionKwargs` dataclass and passes it as a single named kwarg:
```
# nemo_rl/models/automodel/train.py (model_forward)
model_args["flash_attn_kwargs"] = processed_inputs.flash_attn_kwargs
# ...
model(**model_args)
```
This results in the model receiving:
```
**attn_kwargs = {"flash_attn_kwargs": FlashAttentionKwargs(cu_seqlens_q=..., ...)}
```
HF models understand this format — their internal attention layers know to unpack the flash_attn_kwargs dataclass.

AutoModel custom models (GLM4, Qwen3.5) expect flat kwargs:
```
**attn_kwargs = {"cu_seqlens": tensor, "max_seqlen": int, "qkv_format": "thd"}
```
The AutoModel attention layers look for `cu_seqlens` as a top-level key:
```
# nemo_automodel/components/models/deepseek_v3/layers.py:159
cu_seqlens = attn_kwargs.get("cu_seqlens", None) # Returns None!

# nemo_automodel/components/attention/utils.py:92
elif "cu_seqlens" in kwargs: # False! Key is "flash_attn_kwargs"
```
And the THD format trigger is also missed:
```
# glm4_moe_lite/model.py:267
if "qkv_format" in attn_kwargs and attn_kwargs["qkv_format"] == "thd": # False!
# squeeze_input_for_thd() never called
```
Result: TE's DotProductAttention runs with default causal attention on the entire packed tensor, treating all sequences as one continuous sequence.