[HalfSweet]

On Linux, SLogic captures at high sample rates failed after exactly 15 MiB with usb transfer: unknown (errno 12), while macOS was fine. Linux charges the in-flight transfers of every program to one pool, usbcore's usbfs_memory_mb (16 MiB by default). The SLogic driver queued sixteen 1 MiB requests, and the kernel's bookkeeping on top of the sixteenth no longer fit. nusb tried a zero-copy buffer first (the Failed to allocate zero-copy buffer warning), fell back to a heap buffer, and the submission failed with ENOMEM. nusb reports a failed submission in its turn, so the first 15 requests delivered their data and the capture failed on the sixteenth. The sample counts in the report match: 7,864,256 samples × 2 bytes and 15,728,576 × 1 byte are both 15 MiB, less the head bytes. DSLogic USB 3.0 Buffer mode (16 × 1 MiB) and PXLogic at its fastest USB 3.0 rates (four 4 MiB blocks) had the same problem.

Sipeed hit this in its own driver and reworked the transfer ring in libslogic 2.3.3 (sipeed/libsigrok slogic-dev ba92c5fe, sipeed/SLogic f06492f):

• each request holds about 5 ms of data, from 32 KiB to 2 MiB;
• the ring has up to 16 requests;
• at least four must be in flight, because with fewer the analyzer's FIFO overflows and the firmware drops data without reporting it;
• when usbfs refuses a request, the driver takes the ring back before RUN, waits, and halves the request size.

This PR follows that sizing. It plans the ring to fit instead of probing, because nusb reports a refused submission only in its turn, after RUN.

• UsbError::NoMemory: Linux's ENOMEM gets its own variant, and the error key error.device.usbMemory tells the user to raise /sys/module/usbcore/parameters/usbfs_memory_mb instead of showing errno 12.
• UsbTransport::transfer_budget: on Linux, three quarters of the pool, read from sysfs. That is 12 MiB by default, and no limit when the pool is set to 0 (unlimited). Other systems have no budget. Mock devices can set a budget for tests.
• SLogic:
  • requests hold about 5 ms of data, between 32 KiB and 1 MiB (nusb always reads with WinUSB RAW_IO, so 1 MiB stays the largest request);
  • the ring has up to 16 requests within the budget, which is twelve 1 MiB requests by default;
  • in a pool too small for four, the requests shrink instead;
  • if the system still refuses a request, the capture goes on without it while at least four remain, as Sipeed's driver does. The refused request never reached the device, so no data is missing.
• DSLogic and PXLogic: their transfer counts keep within the budget. DSLogic USB 3.0 Buffer mode gets twelve 1 MiB uploads instead of sixteen; PXLogic gets twelve 1 MiB transfers (three blocks) instead of sixteen. Nothing changes on macOS and Windows.
• Docs: the user guide explains Linux's shared pool and what to do if a capture still stops; the SLogic crate README records the later Sipeed revision and how the ring departs from it.

Validation:

• cargo clippy --workspace --all-targets -- -D warnings
• cargo test for hsync-driver, hsync-slogic, hsync-dslogic, hsync-pxlogic, hsync-atklogic and hsync-session: 280 passed. New cases cover: the budget arithmetic; the ENOMEM mapping (the positive case runs only on Linux, i.e. in CI); SLogic plans under the default and small pools; the ring the transport budget allows; captures that go on through refused requests; and the usbfs hint when fewer than four requests remain.
• pnpm test: 86 files, 609 tests passed, including the locale key parity
• Not tested on Linux hardware: the fix is from a user's log and the reference driver; it still needs a capture on a Linux machine with an SLogic16U3 (and ideally a DSLogic U3 in Buffer mode)