Complete reimplementation of the number parser. New parser is 3.5

Complete reimplementation of the number parser.  New parser is 3.5
times as fast as the old one (see data below), and includes the
contribution due to the improved EXACT->INEXACT.

New switch FLONUM-PARSER-FAST? allows the number parser to sacrifice
accuracy for performance, and gains another factor of 2.4 in
performance on flonums.

----------------------------------------------------------------------
Tests performed with a list of 100000 randomly-generated strings.  The
strings were generated by "test-numpar.scm", which implements the R4RS
number BNF, so the strings are arbitrary numeric syntax.

This is a test of general number-reading performance.

Results for old exact->inexact and old string->number:

    process time: 9690 (9060 RUN + 630 GC); real time: 9690
    process time: 9460 (8830 RUN + 630 GC); real time: 9458
    process time: 9450 (8820 RUN + 630 GC); real time: 9451
    process time: 9460 (8830 RUN + 630 GC); real time: 9456

    average real time: 9514 msec

Results for new exact->inexact and new string->number:

    process time: 2800 (2800 RUN + 0 GC); real time: 2800
    process time: 2790 (2790 RUN + 0 GC); real time: 2786
    process time: 2700 (2700 RUN + 0 GC); real time: 2703
    process time: 2680 (2680 RUN + 0 GC); real time: 2686

    average real time: 2744 msec

Average improvement is a factor of 3.5 in speed.