Keep integer true division, hypot() and nextafter() away from the bundled miniexpr until it is fixed #756

To track the follow-up from #754 (#754 (comment)). The bundled miniexpr computes true division, hypot() and nextafter() of integer or bool operands in integer arithmetic (Blosc/miniexpr#26, Blosc/miniexpr#27). #754 gives these expressions NumPy's result dtype, but wherever miniexpr evaluates them the values are still wrong. With #741 and #754:

a = blosc2.asarray(np.array([1, 2, 3, 4]))
b = blosc2.asarray(np.array([2, 4, 4, 4]))
(a / b)[:]                   # [0. 0. 0. 1.]             NumPy: [0.5 0.5 0.75 1.]
(a / b).sum()                # 1.0                       NumPy: 2.75
(a / b + 0.25)[:]            # [0.25 0.25 0.25 1.25]     NumPy: [0.75 0.75 1. 1.25]
blosc2.hypot(a, b)[:]        # [2. 4. 5. 5.]             NumPy: [2.236 4.472 5. 5.657]
(blosc2.hypot(a, b) > 2)[:]  # [False True True True]    NumPy: all True
(a / b).compute()[:]         # right
(a / b)[:2]                  # right

Natively, expr[:] and reductions are wrong. Under WebAssembly, compute() and slices are wrong too; there test_lazyexpr_result_dtype_matches_numpy marks 30 such cases xfail(strict=True) (_WASM_KNOWN_FAILURES). (arctan2() is already handled in python-blosc2.)

Until a fixed miniexpr is bundled, python-blosc2 should send these expressions to an evaluator that gets them right. From the review of #754:

• The guard covers expr[:], compute(), partial slices and reductions, natively and under WebAssembly.
• It reaches an evaluator that is right for these expressions, such as NumPy per chunk. Disabling the prefilter alone may not be enough, if chunk evaluation then picks another miniexpr path.
• It does not convert all inputs to float, which would break integer overflow, comparisons and integers above 2**53.
• It finds the affected operations anywhere in the expression, also below other operators (a / b + 0.25, hypot(a, b) > 2).

When the fixed miniexpr is bundled, the guard goes, and so do the WebAssembly xfails, which then XPASS.

#744 (closed) had a guard for these and other miniexpr bugs (src/blosc2/miniexpr_guard.py), which sent them to numexpr. #741 removes numexpr, so a new guard would send them to NumPy.

AI disclosure: the analysis and the examples were done with Claude Code; I reviewed them.