Wasm: expose building block of compute step in Internal for tests

Context

see #3590 (closed) see #3803 (closed)

To facilitate benchmarking, enough internals are added exposed to be able to decompose one call to compute_step_many into several intermediate steps that can be benchmarked independently.

compute_step_many is refactored a bit to expose enough to be able to

• run on pvm_state rather than tree to measure the impact of encode/decode
• stop the execution sooner than when a snapshotable state is reached, parametrized by a predicate on pvm_state

The internals are exposed in a Internal_for_benchmark module, parallel to Internal_for_tests, to make explicit the contest on which they should be used. A new module (besides _for_tests) has been added to encourage avoiding using them in tests: with them it would be possible to write tests that bypass so much of the main execution loop as to be very fragile.

A module Wasm_pvm_state is introduced to isolate state types declaration.

Remarque: the ppx time-measurement, which provides means to benchmark by putting annotation directly on the code to observe, could allow to obtain some information. However it is out-of-service at the moment (see #3895) and would not allow to isolate parts of the execution.

Manually testing the MR

No new tests, as only internal are made available.

Checklist

• Select suitable reviewers using the Reviewers field below.
• Select as Assignee the next person who should take action on that MR