refactor(optimization): value-only backward

[eric-tramel]

Summary

• Refactor backward pipeline to be Value-only with tape registry + Value.backward entrypoint
• Replace Feedback/Loss/TracedOutput with functional loss fns returning structured Value payloads
• Update execution/training behavior to attach tape ids to Value outputs and propagate through modules/optimizer

Rationale

Aligns plait’s autograd semantics with PyTorch-style loss.backward() while keeping per-record tape behavior and simplifying the API surface.

Changes

• Added tape registry and Value helpers (attach_tape, collect_tape_ids, no_grad, Grad alias) and implemented Value.backward(retain_graph=...)
• Updated executor/module/backward/parameter/optimizer to operate on Value feedback and structured payloads
• Rewrote loss API to pure async functions returning Value (structured list[list[str]] payloads), updated examples/docs/tests
• Removed legacy Feedback, Loss base class, and TracedOutput wrapper

API Notes / Migration

• Training mode now returns Value outputs with tape ids (no TracedOutput)
• Loss fns are async callables returning Value; call await output.backward(grad=loss_val, optimizer=optimizer)
• Parameter feedback accepts Value payloads; structured payload is flattened into strings

Tests

• make ci

Docs

• Updated optimization/inference/execution/parameter docs and examples to new Value-based workflow

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 39d7505a62

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[eric-tramel]

Update: fixed per-record grads in Value.backward (list/tuple and dict support), added unit tests for LossModule and TrainingStep, and added coverage for default optimizer fallback.\n\nTests: make ci; uv run --group dev -m pytest --cov=plait (90% total).