[m-rev.] for review: reduce the parser's memory consumption

[Julien Fischer]

On Fri, 2 Aug 2019, Zoltan Somogyi wrote:

> For tryout and review by Julien. My laptop's timing variability is bigger
> than the performance effect of this change; I hope you can measure its impact
> using Opturion's server class machines.

The output from tools/speed_summary is:

     mercury_compile.asm_fast.gc-new average of 6 with ignore=1      8.84
     mercury_compile.asm_fast.gc-old average of 6 with ignore=1      8.80
     mercury_compile.hlc.gc-new average of 6 with ignore=1      9.06
     mercury_compile.hlc.gc-old average of 6 with ignore=1      8.95

(This above are at -O2.)

Julien.