Do not crash when folding constants that overflow

[Eljees]

Fixes #17008.
Fixes #17534.

These turned out to be two ends of the same pipe, so they are fixed together.

Folding overflows (#17008). constant_fold_binary_int_op and constant_fold_binary_float_op take unbounded ints, and several operations cannot produce a result for large operands. Calling them directly on current master:

int /                        OverflowError: integer division result too large for a float
int << (count > ssize_t)     OverflowError: too many digits in integer
float +  - * / // %          OverflowError: int too large to convert to float
float ** (control)           OK -> None

The ** branches already return None in exactly this situation, so the repository was answering its own question — folding is an optimization, and when it cannot produce a value the expression should simply stand. The other seven paths now do the same.

value_repr (#17534). repr() of an int is bounded by sys.set_int_max_str_digits(), so a LiteralType built from a folded power raised ValueError: Exceeds the limit (4300 digits). It now falls back to hex(), which is lossless and unbounded. This is @devdanzin's suggestion from the thread.

The two are connected: int ** builds the huge value that later explodes in value_repr. Guarding folding alone would leave literals that come from other paths, so both sides are handled.

Verified: the four new tests fail on the unpatched tree with exactly the exceptions above, and pass with the change. mypy/test/testtypes.py goes 119 → 120 passed (2 skipped) with no other movement. black and ruff check clean.

Small operands are covered by a test to pin that folding still happens (6 / 3 == 2.0, 1 << 4 == 16, 2.0 ** 3 == 8.0).

AI-assisted (LLM used for drafting); the reproductions and runs above are mine.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅