Direct rounding for many operators

[obround]

Implements rounding directly using mpfr instead of bigfloat in a number of functions, including all of the major ones. Reduces memory usage on the rival benchmark by 13.6%. Eliminates the use of rnd in most places. I also tried removing the use of parameterize for setting bf-precision (e.g. ival-exact-fabs) like this:

(define (ival-exact-fabs x)
  (define saved-prec (bf-precision))
  (bf-precision (ival-max-prec x))
  (begin0 (ival-fabs x)
    (bf-precision saved-prec)))

However, I was reading an article which suggested using dynamic-wind. We could consider using this to replace parameterize?

[pavpanchekha]

The biggest issue here is the allocation of mutable bigfloat values; we can talk more in meeting today.

A secondary issue is that this is just a huge PR, difficult to review (I skipped reviewing the trig portion). Ideally we'd break it up into pieces.

Finally, I notice that when we use this branch in Herbie runs, the bear benchmark claims "no valid values"; that can't be right and suggests there's a bug somewhere.

[pavpanchekha]

I guess this is fine and all but realistically we should either 1) rewrite the callers to not call these methods, or 2) rewrite these methods into ! variations, in which case the caller decides the precision anyway so these become normal ival-abs! and ival-neg! methods.

[obround]

My understanding is that means replacing something like this:

(define x* (ival-exact-fabs x))

with

(define x* (parameterize ([bf-precision (ival-max-prec x)]) (ival-fabs x)))

This seems unnecessarily verbose to me, or am I missing something?

[pavpanchekha]

My point is that we're often doing ival-exact-fabs in the caller (like hypot) only because we want the max/min logic, not because we actually need to "compute" the fabs. So probably in many cases we can skip actually "computing" fabs (it's just copying bits anyway) if we write the caller in a smarter way.

[pavpanchekha]

Totally fine to leave this PR up if you want to use it as a "cheat sheet" to copy from, but once you're done with it, please close it so we don't just have it hanging around forever. (Even when closed, it's still accessible, so you could even close it now and keep using the PR.)