NEWS:
* Switching endianness mid-byte no longer potentially re-reads bytes.
* bit_offset now consistently refers to the number of bits already
  read.
* HParsedTokens now have a bit_length field; this is a size_t.  This
  may be removed for memory reasons.

The bit writer has not yet been updated to match; the result of
switching bit writer endianness in the middle of a byte remains
undefined.

I've found this especially useful in combination with my own
_attr_uint_const for things like flags and type specifiers.

It's possible that its usefulness might be diminished significantly if
there were a built-in bitfield constant parser -- that certainly would
eliminate all of my current uses of it -- but it still seems nicely
symmetric with H_ACT_APPLY.

Tabs after the first non-tab character are crazymaking. I picked 8
because it wasn't 7 and caused the backslashes to line up on
H_ACT_APPLY.

They apparently removed alloc_size at some point (it was a no-op
beforehand), causing the attribute to throw an error when clang compiles
anything including allocator.h.

It's causing unreferenced-variable warnings, and isn't referenced
anywhere aside from benchmark.c. If client code is likely to reference
it, perhaps move it into another header, so people who include hammer.h
don't have to refer to it to have warning-free code.

NetBSD's character class test macros are defined such that they produce
a compiler warning when given a char argument. Adding a cast to the call
fixes this.

CLOCK_THREAD_CPUTIME_ID is a linux-specific timer. On NetBSD, the best
way to get timer information appears to be getrusage, which happens to
be fairly cross-platform.