feat(provider): route inference via Provider API behind explicit plan override

Introduce dual-transport in CommandCodeLanguageModel: non-Go plans
(goat/pro/max/max20/teampro/provider aliases via normalizePlan) use
the documented Provider API — claude-* → POST /provider/v1/messages
(Anthropic shape), others → POST /provider/v1/chat/completions
(OpenAI shape) with stream:true + Authorization: Bearer + getApiBase()
baseURL. Go/individual-go and absent plan keep legacy
POST /alpha/generate transport byte-for-byte. Both doStream and
doGenerate share providerRunStream with incremental deltas and
terminal usage→finish cost via calculateCommandCodeCost.

Selection order: providerOptions plan → modelOptions plan →
COMMANDCODE_PLAN env, normalized; only go stays legacy. Add
providerBodyFor/providerHeadersFor/providerRunStream with retry/
timeout/abort/redaction parity, extend mock-cc with
/provider/v1/* SSE mocks, and wire provider-transport.test.ts into
test:unit. Refs #51, #53.