Memory management overhaul

[remexre]

On top of #59; merge that one first, then rebase this.

---

• Revamps the memory map to have a single large region of identity-mapped physical memory, and a single large region of RAM. All the previous special regions (beyond the kernel itself) are just part of one of these two.
• Adds a new Devicetree parser that heap-allocates the parsed structure, so the physical memory contained in it can safely be reused.
• Adds a new and much simpler physical memory allocator.
• Adds a virtual memory allocator.
• Adds a heap allocator based on the design from Mimalloc: Free List Sharding in Action.
• Adds some documentation for the above.
  • The mimalloc paper is the "real" documentation for the heap allocator; this code pretty directly follows its design, and divergences are commented on.
  • I'm open to adding more documentation, if there's something that users of the allocators need to know.

---

There are a couple remaining bugs I'm not fixing now:

• Race conditions after freeing from a different hart than allocated memory. The way the delayed free list is described in the paper seems to inherently leave a gap for a race, unless there's extra synchronization not described in the paper.
  • The entire remote free path is TODO()ed out anyway; we've got a bunch of places where we're going to have to make fixes to support multiple harts, so this is just one more for the list.
• Performance improvements; the way randomness is used is slow, and Add a notion of debug builds and debug assertions. #62's debug assertions should be applied to everything in the hot paths of alloc and free.
• A circularity between vma_alloc and alloc. This can probably get resolved by having the kernel virtual allocator pool some memory to use in this case, but I'm going to wait until we have a reproducer before doing this.
• vma_alloc_aligned can spuriously fail when low of virtual address space. Again, I'll wait for a reproducer; this is tricky to trigger in the kernel anyway. Fixing this without a performance regression could be tricky, too; we don't want virtual address space allocation to go from O(log2 n) to O(n) when we have lots of fragmentation, that's a DoS issue.

I'll file these as issues once this is passing review, if none of them seem like blockers to a merge.

---

Resolves #12.
Resolves #27.

[remexre]

Undrafting this; after 1363024, the allocator itself works, and the rest of the changes are definitely reviewable.

This might make sense to split up; if a potential reviewer would like me to do so, I can.

[remexre]

Okay, I think this is fully ready for review and merging.

[jastintime]

started review, devicetree stuff skipped over for now, I need to read up on it prior to reviewing it.

[jastintime]

finished review, src/kernel/devicetree.c and src/kernel/include/devicetree.h aren't reviewed since I don't know enough about devicetree to do so.