perf: improve TFORM CopySpectator performance

[jodavies]

This is a small change to how spectator terms are distributed to TFORM workers. Previously, we ended up creating a huge number of single-term buckets, spoiling efficiency. This now fills the buckets properly.

For a simple benchmark:

#-

#: ScratchSize 500M

Off ThreadStats;
On HumanStats;
On SortVerbose;

Symbol x,n;
CFunction sum;

#ifndef `NTERMS'
	#define NTERMS "{2^25}"
#endif
#define PARTS "{2^6}"
#define TERMSPERPART "{`NTERMS'/`PARTS'}"

Local test = - `NTERMS'*(`NTERMS'+1)/2
	#do p = 0,{`PARTS'-2}
		+ sum(n,{`p'*`TERMSPERPART'+1},{(`p'+1)*`TERMSPERPART'},x^n)
	#enddo
	+ sum(n,{(`PARTS'-1)*`TERMSPERPART'+1},`NTERMS',x^n);
.sort:input-1;
Identify sum(?a) = sum_(?a);
.sort:input-2;

CreateSpectator spec "tmp.`PID_'.spec";

Identify x^n?pos_ = n;
ToSpectator spec;
.sort:to-spec;

CopySpectator zero = spec;
.sort:copy-spec;

Print;
.end

on my usual benchmark machine, with tform -w12, the runtime drops from 32.676 s ± 1.053 s to 13.639 s ± 0.423 s.

The form-bench tests which use spectators don't see any significant improvement.

Benchmark	tform-master -w12 runtime (s)	tform-copyspec -w12 speedup
forcer	39.24 ± 0.18 s	1.01 ± 0.01
forcer-exp	65.29 ± 0.33 s	1.01 ± 0.01
hyperform	32.84 ± 0.21 s	0.99 ± 0.03

The (non-public) forcer-based benchmark from Andreas also sees essentially no change (tentative 1%?)

Benchmark	tform-master -w12 runtime (s)	tform-copyspec -w12 speedup
forcer-av POW=6	61.88 ± 0.35 s	1.01 ± 0.01
forcer-av POW=8	177.11 ± 0.08 s	1.00 ± 0.00
forcer-av POW=10	483.52 ± 1.40 s	1.01 ± 0.00

[coveralls]

coverage: 65.735%. remained the same — jodavies:tform-copyspectator into form-dev:master