[Feature] Adaptive vs Hard Limit on Int32

With PR #167, now use int32 instead of int64 for image wrapper indexing. Currently, all the operators enforce this as a hard limit via the CHECK_TENSOR_INT32_INDEXING validation check, except for CvtColor, which will use int64 indexing if the tensor is too large for int32:

    // line 176 src/op_cvt_color.cpp
    if (needsInt64Wrapper(input) || needsInt64Wrapper(output)) {
        dispatch_cvt_color_itype<int64_t>(stream, input, output, conversionCode, device);
    } else {
        dispatch_cvt_color_itype<int32_t>(stream, input, output, conversionCode, device);
    }

Should figure out in the future whether (some/all) operators should be made adaptive too, or if this limit should be enforced for performance/simplicity.

May also want to document this limit somewhere.

Start by reading PR #167 and tracing the CHECK_TENSOR_INT32_INDEXING validation across the operators. Compare those paths with the adaptive dispatch in src/op_cvt_color.cpp around line 176. Done means the project has an agreed policy for adaptive versus hard limits and the relevant behavior or documentation reflects that decision.