fix(core): make save_adapters/load_adapters round-trip

[SIDDARTHAREDDY8]

Fixes #16102

AdapterModelPTMixin.save_adapters() / load_adapters() did not round-trip. Two independent defects, both in nemo/core/classes/mixins/adapter_mixins.py:

1. A file saved for a subset of adapters could not be loaded with the documented default name=None. load_adapters derived its restore list from the source model's entire adapter config (list(config.keys())) rather than from the file, raising KeyError for every adapter the file never contained. It now derives the list from list(state_dict.keys()), which is by construction exactly the adapters the file holds.

2. Adapter state-dict keys were selected/stripped by substring (adapter_name in k / k.replace(...)), so saving an4 silently serialized all of an4_v2's tensors into the file and broke the subsequent strict load_state_dict. Keys are now matched/stripped on the exact '<adapter_name>.' prefix, so overlapping names no longer contaminate each other.

Test plan

• tests/core/mixins/adapters/ (3 files): 68 passed (TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1), including the 2 new regression tests.
• test_save_adapter_subset_then_load_with_default_name and test_save_load_adapter_with_name_that_is_prefix_of_another both fail on unpatched code (KeyError; foreign keys in saved file) and pass with the fix.

Notes for the reviewer

• This implements the approach from fix(core): make save_adapters/load_adapters round-trip #16103, which the author closed with an invitation to pick it up; the fix hunks and regression tests are carried over from there.
• This change was prepared with AI assistance. DCO sign-off is in the commit message.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.