fix(adk): avoid racing on shared react config when concurrent Run reuses one ChatModelAgent

[GerardGao]

What

buildMessageReActRunFunc and buildAgenticReActRunFunc build the react config
(reactConfig / agenticReactConfig) once per agent and then write the per-run
cancel scope into it on every Run:

msgConf.cancelCtx = cancelCtx
msgConf.modelWrapperConf.cancelContext = cancelCtx

When one long-lived ChatModelAgent is reused across concurrent Run calls
(e.g. a multi-tenant gateway), these writes race with newReact /
newAgenticReact reading the same fields, and two runs can observe each other's
cancel pointer — so a plain run can be cancelled by another run's cancel scope.
Fixes #1177.

How

Inside the run closure, shallow-copy the once-built config before assigning the
cancel fields and hand the copy to newReact / newAgenticReact:

runConf := *msgConf
if msgConf.modelWrapperConf != nil {
    mw := *msgConf.modelWrapperConf
    mw.cancelContext = cancelCtx
    runConf.modelWrapperConf = &mw
}
runConf.cancelCtx = cancelCtx
g, err := newReact(ctx, &runConf)

All other fields are read-only after construction, and buildModelWrappers
snapshots the wrapper config fields at build time, so a struct copy is
sufficient. The per-run cancel injection points (ReAct cancel safe-points, model
wrapper stack, graph interrupt wiring) are preserved unchanged.

Tests

• New TestConcurrentRunSharedAgentWithTools_Race: 4 concurrent runs on one
  shared agent with a start barrier. Fails reliably under go test -race
  before this change and passes after.
• go test -race ./adk/ passes; golangci-lint run ./adk/... reports 0 issues.