fix(subst): convert bound form-locals to expressions lazily

[strub]

Since 615d335 (#1056), f_bind_local / add_flocal eagerly translate
every bound form to an expression so that expression-level locals inside
program statements are substituted as well. Both fix-reduction paths
(EcReduction ι match-fix, EcCallbyValue.try_reduce_fixdef) bind the
constructor arguments through f_bind_local at every recursion step, so
reducing a recursive operator over a large datum re-traverses and
re-hashconses the whole remaining structure at each step: quadratic time
plus heavy GC / weak-table churn. size [0; ...; 4999] = 5000 took ~10s
by cbv delta, simplify delta or done, and ~48s for 10000 elements.

The expression-local maps (fs_eloc, sb_elocal) now hold
expr option Lazy.t: the translation is suspended at bind time and
forced, once per binding, on the first Elocal lookup in e_subst /
subst_expr. None records the CannotTranslate case, preserving the
semantics introduced by #1056.

Also fix a regression from #1056 in EcSubst.rename_flocal: it asserted
that no expression binding existed for the renamed local, but
add_flocal now installs one, so the theory-replay compatibility check
for inductive predicates with renamed constructor binders (e.g.
clone T with pred p <- q) died with an anomaly. The entry is now
overwritten, as f_bind_rename does. Regression test:
tests/clone_ind_rename_binders.ec.

The 5000-element example drops from 10.4s to 0.036s and scales
linearly. tests/eager_call_instanciate.ec, make unit and
make stdlib pass.