The last chunk doesn't need to be created first, instead borrow the
iterator so it can be used later for the remainder.

Make USB endpoint allocator methods accept an optional `EndpointAddress`

chunks_exact() can be handled by the compiler more efficiently.
Previous code was making a memcpy call for each 4 byte chunk slice.
Hoisting the fifo out of the loop avoids recalculating the pointer each
time.

In my benchmark I see a jump from ~13 megabyte/sec to ~25MB/sec after
this change (opt-level=3). opt-level = "z" goes 9MB/s to 18MB/s.

The benchmark was on a stm32h7s3l8, 600mhz clock, 512 byte bulk writes,
data in DTCM. The benchmark isn't just USB writes, also has some
unrelated memcpys for packet construction.

The existing check required N+1 buffer size.

The RM says we have to process STUP (and therefore clear CNAK to start the data stage)
in the DOEPINT STUP interrupt. Seems doing it in RXFLVL when we receive the data is
too early. This makes it work consistently on all chips, so the quirk is no longer needed.

Fixes #3493
Fixes #3459