fix(arange): preserve static output shape for float-dtype arange with integer bounds

[gokulkrishna98]

Problem

torch.arange(start, end, step, dtype=torch.float32) was producing a dynamic-shape output (tensor<?xf32>) even when all three bounds were compile-time integer constants.

Root cause: replace_arange_start_step was casting each operand to the FX node's output dtype (e.g. f32) before calling coreai.range_. When range_ receives float operands it cannot statically determine the element count, so it returns tensor<?xf32> regardless of whether the values are constants.

This propagated through composites that use arange to build position indices (e.g. RoPE), corrupting their output type signature from a static to a dynamic dimension and breaking downstream composites that expected a static shape.

Fix

For integer operands (the common case — e.g. arange(0, head_dim, 2, dtype=float32)): keep all three operands as si32 for the range_ call. The compiler can statically count integer steps, so the result is tensor<NxSI32> with a known N. Cast the result to the requested output dtype afterward.

For float operands (e.g. arange(0.5, 5.0, 0.5)): keep the original behaviour — cast to target_type before range_. A dynamic shape is correct here since floating-point step sizes don't have a statically determinable count, and truncating values like 0.5 → 0 via si32 would produce wrong results.

si32 = IntegerType.get_signed(32)
range_type = si32 if isinstance(start.type.element_type, IntegerType) else target_type

result = coreai.range_(to_scalar(start), to_scalar(end), to_scalar(step))
if result.type.element_type != target_type:
    result = coreai.cast(result, target_type)

Tests

• Updated TestArangeIR::test_dynamic FileCheck patterns to reflect the new IR (si32 operands → range_ → si32 result → cast to f32).
• Added TestArangeIR::test_static_float_dtype_preserves_shape as a direct regression test: arange(0, 8, 2, dtype=float32) must lower to a static tensor<4xf32>, not tensor<?xf32>.
• All existing arange numerical tests continue to pass, including float-operand cases (arange(0.0, 5.0, 0.5), arange(0.5, 5.0, 0.5)).