[BugFix][Codegen] Preserve floating-point literal precision

[anerli]

Problem and root cause

C and Metal code generation use the default stream precision when printing floating-point literals. This can change the value held by a FloatImm: for example, a float32 value near 1.23456789 is emitted as 1.234568e+00f, which rounds to a different float32 value.

Change

• src/target/source/codegen_c.cc: use enough scientific-notation digits to round-trip the stored double value; use a temporary stream for float16 output as well.
• src/target/metal/codegen_metal.cc: apply the same precision while preserving dtype suffixes and existing infinity/NaN handling.
• tests/python/codegen/test_float_literal_precision.py: add device-independent source-generation checks for positive, negative, and small values across C float32/float64 and Metal float32/float16, comparing round-tripped values bitwise.

Tests and validation

12 new regression tests and 6 existing C-host tests pass, 0 skipped. Before the codegen fix, nine new cases fail due to precision loss; the three float16 cases pass.

PYTHONPATH=$PWD/python python -m pytest -q \
  tests/python/codegen/test_float_literal_precision.py
PYTHONPATH=$PWD/python python -m pytest -q \
  tests/python/codegen/test_target_codegen_c_host.py
pre-commit run --files src/target/metal/codegen_metal.cc \
  src/target/source/codegen_c.cc \
  tests/python/codegen/test_float_literal_precision.py

The broader tests/python/codegen/test_target_codegen_metal.py suite segfaults in its first test, test_metal_inf_nan, during execution. The same crash was reproduced after rebuilding unmodified upstream. Thus Metal source generation is validated here, but the Metal execution suite is not green.

Validation environment: Apple M4 Max, macOS/Darwin 24.5.0 arm64, Python 3.12.11. Each branch was rebuilt separately from upstream 907a88c87 with:

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
  -DUSE_METAL=ON -DUSE_LLVM=OFF -DUSE_CUDA=OFF -DUSE_Z3=OFF
cmake --build build --target tvm tvm_compiler tvm_runtime -j 8

git diff --check passes. The applicable formatting and other pre-commit hooks pass. The repository-wide ASF-header hook reports 11 existing violations, reproduced on unmodified upstream; none is in the changed files.

Related

No duplicate open PR was found. This literal-formatting change is independent of Metal target-property discovery in #71.