1 kHz was way too much (the CPU could barely get
anything done in a time window of just 1 ms) and
the task can be put to sleep due to system calls
anyway, so this should be fine.

This is required because the heap is shared among
all tasks and protected using a mutex which only
works in kernel space.

DMA just uses one buffer rather than two now lmao

A relict of debugging on my host.

The allocator works now.
At least when i compile it for amd64.
I haven't actually flashed it yet tho lmao.

The old strategy was to only do context switching
from within the PendSV handler.  This worked fine
until now because all syscalls were handled either
atomically or just returned -EAGAIN if the
resource was locked or busy.  However, with the
introduction of I/O wait, we need to be able to
sleep directly from within the kernel by moving
the context switching completely into the kernel.

It was inevitable, tbh.  I really wanted to do
everything entirely from scratch, but all those
hardware registers are just too much to be
maintained by a single person.  And since i plan
on supporting different boards at some point in
the future, it would be complete madness to redo
everything for that board.