fix: retry openai server errors

[amitksingh1490]

Summary

Treat OpenAI server_error responses as retryable alongside overloaded-server responses, while expanding unit coverage for release, configuration, domain, and retry behavior.

Changes

• Classify both server_is_overloaded and server_error as transient OpenAI failures
• Add release matrix tests for target uniqueness, runner selection, binary naming, cross-compilation, and JSON conversion
• Add coverage for decimal conversion, ordering, serialization rounding, and deserialization
• Add domain tests for file metadata, sync status, model defaults, setters, identifiers, modalities, and serialization
• Add event source retry tests covering exponential, constant, and never-retry policies
• Use pretty_assertions for clearer assertion output in the affected crates

Key Implementation Details

OpenAI retryable error codes are consolidated into a shared constant used by the provider error classifier. Other API error codes, including rate_limit, remain non-retryable through this classification path.

Testing

Exercise the unit tests added or updated in:

• crates/forge_ci/src/release_matrix.rs
• crates/forge_config/src/decimal.rs
• crates/forge_domain/src/file.rs
• crates/forge_domain/src/model.rs
• crates/forge_eventsource/src/retry.rs
• crates/forge_repo/src/provider/retry.rs