Make Matrix implementation more SIMD friendly

This is backporting some changes we made in OSL (with customized
versions of Imath headers that we can thankfully no longer need if we
push the fixes back to the Imath project).

It may be tempting to use memset or memcpy rather than doing certain
element-by-element copies. And indeed, in a scalar context they
generate the same code (with the optimizer on, at least). But when
those operations are done inside a loop that you hope will be
auto-vectorized, the casting and function calling involved in using
memset/memcpy will confuse the vectorizer and you'll end up with
inferior code, sometimes even not vectorizing the whole loop.
This was all pointed out by people from Intel working on OSL, so
I'm deferring to their judgment that this is the best solution.

Signed-off-by: Larry Gritz <[email protected]>