armcc fopen allocated a mutex using the retargeted system-level
_mutex_initialize function. Interestingly, malloc also uses this
same _mutex_initialization function, which prevents a full solution
relying on malloc. The solution previously implemented involved using
the rtx mutex pool for the first 8 mutexes, then falling back on
malloc.

The previous implementation relied on osMutexNew returning an error
on out-of-memory. An unrelated change causes osMutexNew to instead
assert (except for release mode). This meant if you exceed 8 system-
level mutexes in armcc you will hit an assert. Since the filesystem
code can call fopen an unlimited number of times, this is a problem.

Solution is to keep track of which static mutexes we've allocated, so
we know before calling osMutexNew if we need to call malloc.

Also _mutex_free never deallocated the malloced mutexes, which would
cause fopen to leak memory.