fix: tests: buffer the logSource channel

[rowansmithau]

Bug: fakeAgentAPI.PostLogSource sends on an unbuffered logSource channel, which blocks the HTTP handler goroutine if a second PostLogSource call arrives after the test has stopped reading (e.g. from a Kubernetes-spawned pod retrying after a context canceled), causing httptest.Server.Close() to deadlock waiting for the handler to finish.

Fix: Buffer the logSource and logs channels with a capacity of 32, allowing the fake server to absorb all incoming calls without blocking regardless of whether the test is actively reading.

example failure: https://github.com/coder/coder-logstream-kube/actions/runs/27814097460/job/82310947374?pr=201#step:5:472

prior to making changes when testing with:

pass=0; fail=0
for i in $(seq 1 100); do
  if GOTOOLCHAIN=auto go test -count=1 -v -tags=integration \
  -run 'TestIntegration_ReplicaSetEvents|Test_logQueuer/RetryMechanism' \
  -timeout 30s ./... > /tmp/run-$i.log 2>&1; then
    ((pass++)); rm /tmp/run-$i.log
  else
    ((fail++))
  fi
  printf "Run %3d/100 — pass: %d  fail: %d\n" $i $pass $fail
done
printf "\nFinal: %d passed, %d failed out of 100\n" $pass $fail

it would produce a failure rate of 4/100.

post change testing again resulted in 0 failures.

created in Mux with the help of Claude