feat(hooks): populate agent_name on tool-use hook events

[t-mizumoto1203]

What

Tool-use hook events (pre_tool_use / post_tool_use) now carry
agent_name in their hooks.Input, matching every other event type.

Fixes #3109.

Why

hooks.Input.AgentName exists and is documented as "identifies the agent
dispatching the event", and session / turn / before|after_llm_call /
on_agent_switch events all set it. But tool events build their Input via
toolexec.NewHooksInput, which receives no agent, so agent_name shipped
empty — even though the shared dispatchHook path already has the agent.
This blocks any pre_tool_use consumer (auditing, per-agent rate limits,
external policy engines) from knowing which agent issued a tool call.

How

Auto-fill AgentName in dispatchHook (the shared path that already
receives the agent), only when it is empty so explicit caller values win,
mirroring how Executor.Dispatch auto-fills Cwd. The field is omitempty,
so this is backward compatible, and it covers both pre_tool_use and
post_tool_use (both flow through the same path).

Semantics: agent_name = the agent that directly issues the tool call
(a.Name()); on transfer/handoff it follows the active sub-agent.

Testing

• go test ./pkg/runtime/... — added TestPreToolUseHook_ReceivesAgentName;
  all pre_tool_use hook tests pass.
• go build ./..., go vet ./pkg/runtime/..., gofmt — clean.
• End-to-end on a real agent run with a pre_tool_use hook dumping its stdin:
  • single agent: {"agent_name":"root","tool_name":"think"}
  • multi-agent (root delegates to a sub-agent via transfer_task):

    {"agent_name":"root",   "tool_name":"think"}
    {"agent_name":"root",   "tool_name":"transfer_task"}
    {"agent_name":"worker", "tool_name":"think"}
    

    confirming agent_name follows the active agent across delegation.