Fix reciprocal of infinity in turbo loops

[AshtonSBradley]

Summary

Fixes #539 by lowering the literal reciprocal pattern 1 / x to the existing inv compute path instead of the fast division path.

The reported failure happens on AVX-512 when @turbo compiles 1 / Inf through div_fast. That can enter VectorizationBase's pipelined fast division strategy, which mixes division with reciprocal approximation and Newton refinement; the refinement path can turn the mathematically expected zero into NaN for infinite denominators.

This keeps the change intentionally narrow:

• no public API changes
• no new dependencies
• no unrelated refactoring
• only the literal numerator 1 special case is rerouted
• adds a regression test for @turbo y[i] = 1 / x[i] with x .= Inf

Type stability and allocations

For the regression case, the replacement is still type-stable: inv(::Float64) returns Float64, matching the result type of 1 / x[i]. The change happens while constructing the LoopSet, so it should not introduce runtime allocations in the generated loop body. It also reuses the existing inv lowering path rather than adding a new runtime wrapper.

Validation

• formatted touched files with JuliaFormatter using the repository configs
• include("test/testsetup.jl"); include("test/simplemisc.jl") passes from a temporary test environment based on test/Project.toml
• Aqua.test_all(LoopVectorization, ambiguities=false, piracies=false) passes from the same temporary test environment
• focused smoke test for @turbo y[i] = 1 / x[i] with x .= Inf passes
• macro expansion check confirms this pattern now contains :inv and no longer contains :div_fast
• git diff --check passes

I also ran LOOPVECTORIZATION_TEST=part2 julia --project=/private/tmp/lv_test_env test/runtests.jl. It runs with Aqua available and the new issue 539 regression passes, but the shard exits nonzero due to existing ifelsemasks.jl unexpected-pass @test_broken entries unrelated to this change.