fix(qwen): resolve GGUF export failure when MTP config is nested under text_config

[IshantDere]

Closes unslothai/unsloth#8443

Problem

_QwenMtpMixin.__init__ reads num_hidden_layers and mtp_num_hidden_layers
from the top level of self.hparams only. For Qwen3.5/3.6 models these keys
are nested under text_config, so __init__ always read 0 MTP layers. It then
hit assert self.opt_num_mtp_layers != 0, but opt_num_mtp_layers is only
recovered from tensor names during filter_tensors/index_tensors, which run
after __init__. The first GGUF conversion in a fresh process always
crashed with a cryptic AssertionError:

assert self.opt_num_mtp_layers != 0

Fix

1. Flatten config — merge text_config into hparams the same way
   index_tensors already does, so nested mtp_num_hidden_layers /
   num_hidden_layers are found.
2. Replace the bare assert with an actionable ValueError explaining the
   MTP count could not be found and how to work around it
   (--no-mtp, or add mtp_num_hidden_layers to config.json).

Verification

Reproduced against the three relevant config shapes (MTP nested under
text_config, flat top-level config, and a config missing MTP entirely):

Scenario	Before	After
mtp_num_hidden_layers under text_config	AssertionError	25 blocks
flat top-level config	25 blocks	25 blocks
no MTP info anywhere	AssertionError	actionable ValueError

Files changed

• conversion/qwen.py — _QwenMtpMixin.__init__