refactor(models): (PR 1/4) isolate plugin registration failures

[Luodian]

This makes one plugin payload register transactionally and keeps later entry points visible when an earlier one fails. Built-in resolution and lazy backend imports remain unchanged.

Depends on #1440.

GitHub review stack #1468.

Stack

• 1/4 refactor(models): (PR 1/4) isolate plugin registration failures #1446 (this PR) - transactional plugin registration
• 2/4 refactor(models): (PR 2/4) derive legacy views from V2 #1447 - read-only legacy views
• 3/4 refactor(models): (PR 3/4) route model discovery through V2 #1448 - V2 discovery consumers
• 4/4 docs(models): (PR 4/4) align registry contribution guidance #1449 - contribution guidance
• Merge surface: refactor(models): (collector) make ModelRegistryV2 authoritative #1450 collector only

This is review layer 1 of 4. Do not merge this PR; the final collector will land the complete authoritative registry atomically.

Summary

• Register one or many manifests through copy-then-commit state.
• Return structured failures per entry point and continue loading later plugins.
• Preserve strict conflicts and explicit overwrite behavior.

In scope

• ModelRegistryV2 transaction, listing, and entry-point failure behavior.
• Focused rollback, conflict, and isolation tests.

Out of scope

• Global startup, compatibility views, and discovery consumers are later review layers.
• This layer is not an independently mergeable release unit.

Validation

• python -m pytest -q test/models/test_model_registry_v2.py | sample size: N=18 tests | key metrics: 18 passed | result: pass
• pre-commit run --files lmms_eval/models/registry_v2.py test/models/test_model_registry_v2.py | sample size: N=2 files | result: pass

Risk / Compatibility

• Built-in names, aliases, chat precedence, and force_simple remain unchanged.
• The collector is the only branch authorized to merge this registry stack.

Type of Change

• Bug fix (non-breaking change)
• New feature
• New benchmark/task
• New model integration
• Breaking change
• Documentation update
• Refactoring (no functional changes)