refactor(recipes): add typed tokenizer construction

[akoumpa]

What

• add TokenizerConfig.build() for tokenizer and multimodal processor factories
• resolve dataset overrides, explicit null, top-level factories, model fallback, and trust policy at the RecipeConfig boundary
• remove the private, signature-inspecting _build_tokenizer helper and its cross-recipe import
• move LLM, sequence-classification, and retrieval recipes onto the typed build API
• replace retrieval processor type dispatch with capability-based tokenizer access

Why

This is the second follow-up layer on #2390 and depends on #3008. It keeps declarative tokenizer construction out of recipe bodies and preserves config state across builds.

Validation

• focused stack suite: 159 passed, 7 skipped
• tokenizer boundary suite: 6 passed
• uv run ruff format .
• uv run ruff check --fix .
• uv run lint-imports
• uv run python -m compileall -q nemo_automodel

Stack

• root: refactor(datasets): typed Config + build per dataset, config-driven dataloader #2390
• base: refactor(recipes): move LLM packing policy into config #3008
• this PR

Checklist

• Followed the contributor and repository review guidance
• Added focused tests
• No documentation change is required for this internal construction refactor

[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.

[akoumpa]

/ok to test 90f69ef