Metal: express the intrinsic lowering as tables and small helpers

[maleadt]

lower_llvm_intrinsics! handled every intrinsic in one loop body, repeating the declaration lookup, builder setup and type mangling for each kind of intrinsic. This expresses the lowering as tables plus a few small helpers:

• AIR_MATH_INTRINSICS, AIR_VALUE_INTRINSICS, AIR_BIT_INTRINSICS and REMOVABLE_INTRINSICS map LLVM intrinsics to their AIR functions.
• declare! gets or creates a declaration, and checks that an existing one has the expected type; call_declared! and type_suffix cover the call and the mangling.
• lower_intrinsic_calls! collects the intrinsic calls of a function, then replaces each by what a lowering function returns (a value, :erase, or nothing to keep the call). It doesn't iterate to a fixed point or erase anything else.
• The minimum/maximum wrapper, copysign and is.fpclass lowerings become separate functions. fuse_minmax3! keeps its fold-and-rescan loop.

The section shrinks from 435 to 334 lines of code.

One behavior change: a declaration that already exists with a different type is now an error. Before, the lowering silently called it with the wrong type.

Verification

• The generated code is unchanged: on a corpus of intrinsic calls (the math, value and bit intrinsics, minimum/maximum, copysign and is.fpclass, in scalar, vector and bfloat variants, with and without fast-math flags, and chained min/max), the output of lower_llvm_intrinsics! and of the full lower_air! is byte-identical to main for Metal 3.2/AIR 2.7, 4.0/2.8, 4.0/2.9 and 4.1/2.9.
• It's also a bit faster: on a module with 19,000 instructions of intrinsic calls, lower_llvm_intrinsics! takes 60 ms instead of 87 ms. On a 200,000-instruction module without calls it's on par (100 ms vs 102 ms).
• GPUCompiler's metal tests pass on Julia 1.12 and 1.10.
• Metal.jl's full test suite passes against this branch, except for examples/flopscomp, which fails because BFloat16s isn't installed in the test environment.

[codecov]

Codecov Report

❌ Patch coverage is 92.30769% with 18 lines in your changes missing coverage. Please review.
✅ Project coverage is 86.42%. Comparing base (70ad652) to head (829bc12).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/metal.jl	92.30%	18 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #941      +/-   ##
==========================================
+ Coverage   86.32%   86.42%   +0.09%     
==========================================
  Files          29       29              
  Lines        5756     5672      -84     
==========================================
- Hits         4969     4902      -67     
+ Misses        787      770      -17     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.