fix(test): stop the backend-trace specs racing the lossy trace channel

[localai-bot]

Problem

The tests workflow failed on master at 0a8a7fbb (run 30244703942):

[FAILED] Timed out after 1.000s.
Expected <int>: 158 to equal <int>: 200
In [BeforeEach] at core/http/endpoints/localai/traces_test.go:120

Found by the pre-release CI sweep for v4.8.0. The spec was added in #11056, so it is new in this release cycle.

Cause

This is not a slow test, and raising the timeout would not have fixed it.

RecordBackendTrace does a non-blocking send onto backendLogChan, which is buffered at 100, and drops on the default: branch when full:

select {
case backendLogChan <- &t:
default:
    xlog.Warn("Backend trace channel full, dropping trace")
}

That is correct and deliberate: tracing must never stall inference. But the spec pushed 200 traces of ~16 KB in one tight loop. When the consumer goroutine does not keep up, entries are dropped permanently, and Eventually(...).Should(Equal(200)) can never be satisfied no matter how long it waits. On an idle machine the consumer keeps up and the spec passes, which is why this only shows up under CI load.

Fix

Feed the traces in chunks of 50, draining after each, so the channel is never overrun and the count stays exact. Production code is unchanged: the lossy fast path is the right behavior, the test was wrong to depend on it.

Verification

Reproduced and fixed under identical synthetic load (60 busy loops on a 20-core box, GOMAXPROCS=2), 12 runs each:

	Result	Observed counts
Before	0 / 12 passed	stuck at 164, 101 (CI saw 158)
After	12 / 12 passed	200 every run

Full package also green: go test ./core/http/endpoints/localai/ → ok (34.9s).