Fix output type of custom calls while lowering quant/dequant torch op to HLO

[sdasgup3]

#5763 allows lowering the torch quantize/dequantize operations to HLO custom calls. For example,
the following PyTorch code

x = torch.ops.quantized_decomposed.quantize_per_tensor(
    x, 0.4, 2, -128, 127, torch.int8)
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
    x, 0.4, 2, -128, 127, torch.int8)

is lowered to the following HLO operations:

ENTRY %IrToHlo.5 (p0.1: f32[2,3,4,5]) -> (f32[2,3,4,5]) {
  %p0.1 = f32[2,3,4,5]{3,2,1,0} parameter(0)

  %custom-call.2 = f32[2,3,4,5]{3,2,1,0} custom-call(f32[2,3,4,5]{3,2,1,0} %p0.1), custom_call_target="stablehlo.uniform_quantize", api_version=API_VERSION_TYPED_FFI, backend_config={scale=[0.4],zero_point=[2],storage_type=si8,expressed_type=f32,storage_min=-128,storage_max=127}

  %custom-call.3 = f32[2,3,4,5]{3,2,1,0} custom-call(f32[2,3,4,5]{3,2,1,0} %custom-call.2), custom_call_target="stablehlo.uniform_dequantize", api_version=API_VERSION_TYPED_FFI, backend_config={scale=[0.4],zero_point=[2],storage_type=si8,expressed_type=f32,storage_min=-128,storage_max=127}
  ROOT %tuple.4 = (f32[2,3,4,5]{3,2,1,0}) tuple(f32[2,3,4,5]{3,2,1,0} %custom-call.3)
}

Note that the output of custom call corresponding to quantize op has element type f32. The fact that the output of custom_call (for quantize operation) is a float is more of a logical problem as the the result of quantization is generally expected to be in integer domain. Also, note that the choice of output type should not effect the eventual conversion of HLO custom calls to mhlo uniform.quantize/uniform.dequantize operations.

Moreover, based on https://github.com/pytorch/pytorch/blob/0b72ce1bd1a4a0596dde4053899b8a9a7999bc47/torch/ao/quantization/fx/_decomposed.py#L164 we set the output element type of dequatize operation to be f32

Finally improved the debuggability of the map queries using proper error messages.

cc @GleasonK

[GleasonK]

We could implement this as a switch / series of if stmts to avoid a static dictionary, it's a short enough list. Not sure of PT/XLAs stance on static data.

[sdasgup3]

I agree about less static data floating around. Replaced the dictionary with conditionals.
@lsy323 If the change LG to you then I can merge once the CI is green.