feat(anthropic): add prompt caching support (memohai#9)

* feat(anthropic): add prompt caching support

* fix(ci): golang lint

Release v0.3.4

sdk: remove unused MaxTokens field from Model struct

The field was never read anywhere; actual max-tokens control is handled
at the request level via GenerateParams.MaxTokens.

fix(anthropic): set default max_tokens when unset (memohai#1)

fix(sdk): harden step loop cancellation and parallelize tool execution

- Add ctx.Err() checks at step loop entry in both StreamText and
  GenerateText to bail out immediately when the context is cancelled,
  instead of attempting another DoStream/DoGenerate call.
- Add a second ctx.Err() check in StreamText after the inner stream
  consumption loop exits, so a cancelled context skips step result
  processing and tool execution.
- Refactor executeTools into two phases: sequential approval checking
  followed by parallel tool.Execute via goroutines, using indexed
  result slots to avoid races. Single-tool calls skip goroutine
  overhead.

docs: update MCP

fix: remove duplicate /v1 prefix from Anthropic ListModels/Test/TestM…

…odel paths

defaultBaseURL already includes /v1, so API paths should not repeat it.
Also fixes pre-existing TestDoGenerate path assertion.

fix: resolve all golangci-lint issues across codebase

- errcheck: check json.Unmarshal return values
- goconst: extract repeated string literals into constants
- gocritic: pass large structs by pointer, fix rangeValCopy,
  ifElseChain, evalOrder, typeSwitchVar
- prealloc: preallocate slices where possible
- staticcheck: use type switch assignment, struct conversion
- unused: remove unused type