[mlir][dataflow] IntRange: Replace yield-based widening with per-state lattice budget (#196616)

IntegerRangeAnalysis can hang on `scf.while` loops with dynamic bounds:
a
loop-carried range ratchets [0,0]->[0,1]->[0,2]->... by one per worklist
visit, requiring up to 2^31 iterations on i32. The new
`int-range-analysis-convergence.mlir` test reproduces this.

The ratchet lives at framework merge sites (region successors, callable
args) where the solver joins lattices via virtual
`Lattice::join(const AbstractSparseLattice &)`. The pre-existing
`isYieldedResult`/`isYieldedValue` heuristic in
`IntegerRangeAnalysis::visitOperation` doesn't help: it runs in the
transfer-function callback for inferrable-op results used by a
terminator,
not on the merge path. It is also harmful where it fires - slams to
maxRange on the *second* visit (after, say, [1,1]->[1,2]), so naturally
bounded accumulators (e.g. `arith.minsi`-clamped iter args) widen to
[INT_MIN, INT_MAX].

Replace it with a per-state widening budget on
`IntegerValueRangeLattice`:
the lattice counts merge-site joins and forces the range to its max once
the count hits `kIntegerRangeWideningBudget` (128). Only the virtual
overload is overridden, so transfer-function joins via the non-virtual
`join(const ValueT &)` are unaffected. The new
`int-range-loop-iter-args.mlir`
test pins the tighter bounds; the convergence test verifies termination.