fix delayed harness tool lifecycle events

Finding: In @ai-sdk/harness, when a HarnessAgent runtime starts multiple host-tool calls in one step, onToolExecutionStart and onToolExecutionEnd were reported to be delivered only at finish-step, back to back after tool output already existed, and a fast tool's result waited for a slower one. Source: https://github.com/vercel/ai/issues/21685. This is an external report with no KFA reproduction. Version and environment: The issue reports @ai-sdk/harness 1.0.128 and notes the repository version at filing was 1.0.132. A contributor comment says 1.0.132 also reproduced the behavior on main, so upgrading from 1.0.128 to 1.0.132 did not help (https://github.com/vercel/ai/issues/21685#issuecomment-5885957867). The runtime and operating environment are not specified. Reported reproduction: Use a HarnessAgent with a runtime that starts two host-tool calls in one step, one fast and one deliberately slow. Register both callbacks and log execute start/end alongside callback and tool-result arrival. The issue reports callbacks only at the step boundary (https://github.com/vercel/ai/issues/21685). The issue says this contradicts the documented contract that onToolExecutionStart runs before execute. Expected behavior: start before each execution, and end plus result as each tool completes, with finish-step still marking the model-step boundary. Reported cause: A comment attributes the problem to step accounting and real-time tool execution sharing a single publication buffer (https://github.com/vercel/ai/issues/21685#issuecomment-5886266359). This is a reported statement; the diff was not inspected, so treat it as source-reported rather than independently confirmed. Fix status: The same comment says PR #21696 publishes lifecycle events and results per tool while keeping finish-step as the step boundary, and lists passing tests. A later comment lists @ai-sdk/harness 1.0.133 as published (https://github.com/vercel/ai/issues/21685#issuecomment-5887817259). Report, PR, and release are separate claims; we did not verify the release contents or run the fix. Diagnosis hint, source-backed: if your symptom matches (callbacks and fast results delayed until finish-step with parallel host tools) and you use a version at or below 1.0.132, the reported fix is in 1.0.133 or later. No workaround was documented. Claude Code-specific progress forwarding is tracked in #21686 and is out of scope. Limitations: contributor-association comments, automated-looking format, no inspected diff, no KFA reproduction.