[fjtrujy]

Scope

This PR isolates PSP asynchronous USB receive lifecycle and cleanup from the earlier combined #72.

• Process detach before an async receive completion and do not submit requests when the connection is gone.
• Handle successful zero-length completions by rearming the receive endpoint, without treating a zero-length USB packet as an error.
• Treat retcode == -3 as cancellation (per PSPSDK), not as a transient I/O error; clear the connected state and wait for the usual attach/HELLO handshake instead of resubmitting on a cancelled endpoint.
• For other receive errors, schedule a retry through a 100 ms timed event wait rather than blocking the USB event thread with sceKernelDelayThread. Detach/attach events remain serviceable while waiting.
• Retry a failed sceUsbbdReqRecv() submission while the USB bus remains connected, rather than logging once and permanently losing the receiver.
• Validate the receive length before copying data and check sceUsbGetState() & PSP_USB_CONNECTION_ESTABLISHED before queueing new work.
• Only mark connected after the initial async receive has been queued successfully; clear connection state on start/stop and fix the event flag cleanup assignment.

Verification

• make -C usbhostfs -j4 with the installed PSP SDK: passes.
• make -C tests test: existing host C regression passes.
• git diff --check: passes.
• Rebased onto current master so already merged directory, host handle and PSP bulk-transfer fixes are not reverted or duplicated.

Remaining validation

The revised cancelled/error/zero-length completion logic has not been tested on a physical PSP yet. In particular we need to exercise USB disconnect/reconnect, error and cancellation behavior. Build success and a host-side test do not validate the PSP's kernel USB request lifecycle.

The independently investigated long-idle hang remains tracked in #74; this PR does not claim to fix its root cause. Please re-review the revised async error handling before merging.