fix(llm): gate OpenAI/Compatible supports_vision on model capability

[bug-ops]

Summary

• OpenAiProvider::supports_vision() unconditionally returned true regardless of the configured model, so a text-only model (e.g. gpt-3.5-turbo, or an arbitrary text-only type = "compatible" endpoint) would have images attached and sent, typically failing with a 400. CompatibleProvider inherited the same bug via delegation.
• supports_vision() now resolves via an explicit vision override (new OpenAiConfig/CompatibleConfig field, with_vision builders, and a new vision field on [[llm.providers]] entries) falling back to a built-in model-name prefix table for well-known OpenAI vision families (gpt-4o, gpt-4-turbo, gpt-4-vision, gpt-4.1, gpt-5, vision-capable o-series), explicitly excluding text-only reasoning variants (o1-mini, o1-preview, o3-mini). Unrecognised model names fail safe to false — this is also the default for arbitrary compatible endpoints unless vision = true is set explicitly.
• Same defect class as zeph-llm: Ollama provider's supports_vision() hardcoded true regardless of configured model #6377, already fixed for Ollama in fix(llm): gate Ollama supports_vision on actual model capability #6410.

Closes #6411

Test plan

• Full local check suite: cargo +nightly fmt --check, cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings, cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins (14217 passed, 0 failed, 35 skipped; one unrelated concurrency test flake confirmed non-reproducing on re-run), rustdoc gate with RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" — all clean.
• cargo test --doc --workspace — all 32 crates pass, including new with_vision doctests on OpenAiProvider/CompatibleProvider.
• Per-crate rustdoc check for touched crates (zeph-llm, zeph-config) — clean.
• New unit tests cover: known-vision-model auto-detect, gpt-3.5* false, unrecognised-model fail-safe false, explicit override in both directions, text-only o-series exclusion (o1-mini/o1-preview/o3-mini), CompatibleConfig.vision config-field forwarding (distinct from the .with_vision() builder path).
• LLM serialization gate: live round-trip against a real gpt-4o-mini OpenAI provider confirmed the request reaches the server and is rejected only with 429 (rate-limited, a known pre-existing test-account constraint, unrelated to this change) — never a 400/422, confirming the request payload is well-formed. Verified in source that 400 responses are routed through a distinct error path (map_error_response → LlmError::ApiError) never conflated with the 429/503 retry loop that was actually hit.
• Adversarial critique (impl-critic) and independent test-coverage validation (tester) both ran; one significant finding (o-series text-only false positive) was fixed and re-verified before code review approval.