Return errors for undelivered HID reports

[BenDManning]

Stacked on #1516 so its recovery work stays the starting point. This is not a competing PR to dev.

The additional contract needed by BenDManning/jetkvm-mcp#146 is synchronous: a report may only be acknowledged after its HID endpoint accepts the complete write. Later recovery cannot turn an already dropped report into success.

Changes:

• require USB state configured before attempting keyboard, absolute mouse, relative mouse, or wheel reports;
• return temporary HID errors and the first write timeout instead of suppressing them;
• reject reports sent to disabled keyboard or mouse endpoints;
• treat short writes as failures;
• retain only successfully written keyboard state, preventing a rejected keyboard report from being replayed later by the recovery probe;
• preserve Escalate USB recovery when keyboard HID writes time out #1516 recovery and make its timeout streak observe every HID endpoint, even after the failed file handle closes.

Tests deterministically fill a pipe to stall each endpoint and verify the first report returns os.ErrDeadlineExceeded. Companion cases read the exact bytes for every successful/acknowledged report. The common RPC/gadget seam verifies rejection for every non-configured USB state and exact write-error propagation.

Local checks:

• GOMODCACHE=/tmp/jetkvm-kvm-146-gomod GOCACHE=/tmp/jetkvm-kvm-146-gocache go test -race ./internal/usbgadget -count=1
• GOMODCACHE=/tmp/jetkvm-kvm-146-gomod GOCACHE=/tmp/jetkvm-kvm-146-gocache go vet ./internal/usbgadget
• git diff --check

Independent review at exact head 2408a67f23a69f04c2758588f01ee836bf7222fb:

• Standards: 0 findings after the duplicated-test-case finding was resolved
• Spec: 0 findings; physical/post-install gates remain intentionally outstanding

Upstream checks at that exact head:

• Build and test: pass
• Go: pass
• UI: pass
• CLA: pass

No firmware was installed or flashed, and no device mutation occurred.

[CLAassistant]

All committers have signed the CLA.