feat: add Tsubasa provider configuration

[cenab]

Summary

Add Tsubasa to the existing provider configuration and README setup list. The named entry resolves TSUBASA_API_KEY, offers tsubasa-pro and tsubasa-fast, and reuses OpenAI Chat Completions. Both models have 32,768-token context metadata and text-only input; tool, parallel-tool, and reasoning capability flags remain false.

No Rust runtime, authentication implementation, dependency, price, release date, or underlying-model identity is added.

Verification

• cargo check -p forge_repo: passed.
• cargo test -p forge_repo --lib provider::provider_repo::tests: 17 passed.
• Full forge_repo library harness: 333 passed, 1 existing ignored test. Built with cargo test -p forge_repo --lib --no-run --message-format=json, then executed with isolated FORGE_CONFIG, preserved HOME, synthetic/empty credentials, loopback-only networking, and the supported INSTA_WORKSPACE_ROOT override. The first harness attempt could not run Cargo metadata under its personal-file restrictions; correcting snapshot-root configuration resolved those harness failures. Serial test execution also avoided contention in an unrelated heartbeat timing test.
• cargo fmt --all -- --check: reports one unchanged upstream file, crates/forge_main/src/info.rs, which is byte-identical to base 574894e8e9ab191788d451d569eb15bfcf55ed74. This change adds no Rust source. JSON is parsed by the actual provider loader; git diff --check passed.
• Actual ForgeInfra → ForgeRepo → ForgeServices provider selection, environment-to-credential migration, template rendering, production request transformers, and HTTP/SSE decoded Hello for both aliases against a controlled loopback endpoint. Invalid credentials returned 401. Without TSUBASA_API_KEY, Tsubasa remained unconfigured despite unrelated OpenAI and OpenRouter keys.
• All three captured HTTP request bodies passed Tsubasa's real strict schema and conservative context/output preflight. Only the resolved provider URL was redirected to loopback; the transport, parser, and credential handling were unmodified.

This validates the local provider path, not a full interactive Forge session or live tool quality. The latest public Tsubasa /v1/models check returned HTTP 404; live inference qualification remains outstanding. The fixture used synthetic credentials and controlled replies.

Prepared with AI assistance and submitted as a draft for maintainer review.

Co-Authored-By: ForgeCode noreply@forgecode.dev

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}