[KernelGen] Add view_as_complex operator

[zacliu2023]

Summary

Add view_as_complex operator for Iluvatar (Tianshu) platform that converts real tensor (..., 2) to complex tensor (...).

Generated with kernelgen MCP v2.0 and validated on Iluvatar CoreX BI-V150 hardware.

Implementation Details

• Platform: Iluvatar (Tianshu) CoreX BI-V150
• Function: Convert last dimension 2 of float32/float64 tensor to complex64/complex128
• Approach: Use PyTorch native implementation which is highly optimized

Test Results

Accuracy Tests

dtype	Shape	Result
float32	(4, 2)	PASS
float64	(8, 16, 2)	PASS
float32	(1024, 1024, 2)	PASS
float32	(0, 2)	PASS (empty)

Features

• ✅ float32 input -> complex64 output
• ✅ float64 input -> complex128 output
• ✅ Multi-dimensional tensor support
• ✅ Empty tensor handling
• ✅ Proper error handling for wrong input shape

Files Changed

• src/flag_gems/runtime/backend/_iluvatar/ops/view_as_complex.py - view_as_complex implementation
• src/flag_gems/runtime/backend/_iluvatar/ops/__init__.py - Operator registration

Testing Commands

python3 -c "
import torch
from flag_gems.runtime.backend._iluvatar.ops.view_as_complex import view_as_complex
input = torch.randn(4, 2, dtype=torch.float32, device=\"cuda\")
output = view_as_complex(input)
print(f\"Input: {input.shape} -> Output: {output.shape}, {output.dtype}\")
"

Checklist

• Code follows FlagGems coding standards
• All accuracy tests pass
• Operators registered in backend __init__.py
• Generated with kernelgen MCP v2.0