use arith 'exact'/'fast'

[rurban]

New lexical pragma for exact or fast arith, promote to double or bigint? We probably need a PL_hints bit for this, more then for no magic or strict names.

Add a core pragma to select arithmetic overflow semantics.
Overflow to fast NV (double) as with perl5, or to exact bigint/num/rat as with perl6.

[rurban]

See e.g. 2713d58
for pp_add with exactarith instead of promoting to NV. Note that NV overflow/precision loss is not done there yet.