Reduce allocations in builtin range

[vdpoora]

range() builds its result array with one []Int backing slice and a
presized []Object, so construction takes a small constant number of
allocations regardless of length, instead of one boxed Int per element plus
repeated slice growth.

It also computes the element count with overflow-safe unsigned math, and caps
materialization at a new MaxRangeLen (default 10,000,000; at ~24 bytes per element that bounds one
range at ~240 MB), returning ErrRangeLimit instead of a makeslice panic on an
oversized range.

Benchmarks

go test -bench . -benchmem, measured per Compiled.Run, this branch vs master:

scenario	master	this PR
x := range(0, 1000000)	1000041 allocs / 96 MB	5 / 24 MB
for i in range(0, 100000) { s += i }	200031 / 10.6 MB	100005 / 3.3 MB
for _ in range(0, 1000000) {}	1000042 / 96 MB	6 / 24 MB

Construction drops from ~N allocations to a small constant at any length. The
allocations left in value-using loops are the loop-body arithmetic, not range
construction.

Notes

• range() still returns a materialized array; value semantics are unchanged.
• The overflow-safe count also fixes wrong results for wide ranges, e.g.
  range(0, 10, 9223372036854775806) now yields [0] instead of an empty
  result.
• The MaxRangeLen cap is a behavior change at extreme sizes (an oversized
  range() returns an error instead of OOM/panic). Happy to split it into its
  own commit if you'd prefer to review it separately.