Extract primitive binary operations from generic derivation

[craigfe]

This PR is just code movement, in preparation for a follow-up that improves the
implementation of Repr.size_of.

[craigfe]

I wasn't sure if this code move would be a perf regression due to inlining differences, but it seems not. Taking a random subset of the benchmarks on 4.12+flambda:

/bin/bool/decoder

  major-allocated (words) 1.912643276295168e-08 -> 1.8749617157615887e-08
  minor-allocated (words) 3.000144818917675     -> 3.00014338458947
  monotonic-clock (ns)    14.364530138134555    -> 14.2387131952349

/bin/bool/encoder

  major-allocated (words) 0.00013272412368469736 -> 0.0001325673753334568
  minor-allocated (words) 17.000369171136818     -> 17.000365513605104
  monotonic-clock (ns)    36.614114062016235     -> 36.202986100166434

/bin/int/char (0 <= b <= 7)/decoder

  major-allocated (words) 2.2426733518418638e-08 -> 2.333720880839565e-08
  minor-allocated (words) 3.00015682243577       -> 3.000159975708209
  monotonic-clock (ns)    15.490271721869108     -> 15.827316030885317

/bin/int/char (0 <= b <= 7)/encoder

  major-allocated (words) 0.00011094330024032022 -> 0.00011439560427351727
  minor-allocated (words) 13.988809511338692     -> 14.000392920751212
  monotonic-clock (ns)    46.64706870512427      -> 45.84583076612596

/bin/triple_short_int/decoder

  major-allocated (words) 0.0003661506530433007 -> 0.0003661506530433007
  minor-allocated (words) 16.00026315715933     -> 16.00026315715933
  monotonic-clock (ns)    26.0765536583521      -> 25.954999735241444

/bin/triple_short_int/encoder

  major-allocated (words) 0.00047488916106462863 -> 0.00047772855653753846
  minor-allocated (words) 20.001231993250016     -> 20.001244328326532
  monotonic-clock (ns)    121.5038156944062      -> 123.64678945748766

/bin/variant/decoder

  major-allocated (words) 0.00022407974378948177 -> 0.0002068110465555966
  minor-allocated (words) 22.350997232953027     -> 22.323529632453084
  monotonic-clock (ns)    37.476367788729206     -> 38.48230385789799

/bin/variant/encoder

  major-allocated (words) 0.001118561664162485 -> 0.0011464753845754887
  minor-allocated (words) 31.659976389036725   -> 31.671412651120534
  monotonic-clock (ns)    89.58312741177834    -> 94.87063236576186

These differences look to be within measurement noise to me.

[mattiasdrp]

Love this PR, the code looks so much cleaner, thanks!

[mattiasdrp]

Let's merge this fast to reduce the noise for the important PR ;-)