…i-activation

Adds explicit fan-in support to the workflow engine while preserving
adk-python's semantics for ordinary nodes (multiple incoming edges
on a non-Join node still fire it once per upstream completion).

Engine changes:
* Scheduler keys per-task state by a monotonically increasing
  activationID instead of node name, so a node can have multiple
  concurrent activations without overwriting each other's bookkeeping.
* Per-name FIFO trigger buffer serialises activations of the same
  node, so JoinNode (and any custom WaitForOutput=true node) sees
  per-predecessor activations one at a time and can safely
  accumulate state.
* WaitForOutput honoured in handleCompletion: a completed activation
  that produced no "output" event lands in NodeWaiting and skips
  successor scheduling, deferring the join to the next predecessor.
* Per-(run,name) joinAccumulator created lazily on first activation
  and cleared once the fan-in node emits its terminal output.

API additions:
* workflow.JoinNode + workflow.NewJoinNode: aggregates per-predecessor
  inputs into a single map[string]any output once every predecessor
  declared in the graph has fired. Forces Config().WaitForOutput=true
  regardless of caller input.
* (*nodeContext).InNodes() exposes the static predecessor-name set
  derived from the graph at construction; mirrors adk-python's
  Context.in_nodes.
* graph.inNodeNamesOf for engine-internal predecessor lookup.

Validation:
* validateJoinNodesHaveIncoming + ErrJoinNodeNoIncoming reject
  graphs whose JoinNode has zero incoming edges. adk-python defers
  this to runtime; we surface it at workflow.New so misconstructed
  graphs fail to build.

Tests:
* TestScheduler_MultiActivation_NonJoinNode_RunsTwice locks in the
  Python-parity "no automatic merge" behaviour for ordinary nodes.
* TestScheduler_WaitForOutput_{StaysWaitingWithoutOutput,
  EmittingOutputSchedulesSuccessors} cover both branches of the new
  WaitForOutput path.
* TestScheduler_TriggerBuffer_SerialisesSameNameActivations asserts
  the per-name FIFO contract JoinNode relies on.
* TestJoinNode_E2E_FanIn{Two,Three}Branches exercise the canonical
  fan-in graph shapes from the API design doc.
* TestJoinNode_Run_DetachedYieldsDegenerateOutput keeps the node
  usable in unit tests that exercise Run() outside the engine.
* TestJoinNode_Run_RejectsUnknownPredecessor + the validation tests
  cover the error paths.

All existing tests still pass; the suite is race-free under
go test -race.