Update the report output to be a bit cleaner. I've made the following
changes:
 * Round decimal values to 3 decimal places.
* Remove mean signed deviation and L2 norm. We have SNR, mean absolute
error, and max element-wise error. I think that's sufficient.
* Rename SQNR to SNR. Not all tests are quantized and we're using it as
a general measure of tensor-wise error.
* Split out the result column into pass/skip/fail and a detailed reason.
This should be easier to parse at a glance.

Example output (not sure markdown formatting is weird here...):

Test ID | Test Case | Flow | Params | Result | Result Detail | Delegated
| Quantize Time (s) | Lower Time (s) | Delegated Nodes | Undelegated
Nodes | Delegated Ops | Undelegated Ops | PTE Size (Kb) | Output 0 Error
Max | Output 0 Error MAE | Output 0 SNR | Output 1 Error Max | Output 1
Error MAE | Output 1 SNR
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
| -- | -- | -- | -- | --
test_add_dtype_float32_xnnpack | test_add_dtype | xnnpack | {'dtype':
torch.float32} | Pass |   | TRUE |   | 3.747 | 1 | 0 |
{'aten::add.Tensor': 1} | {} | 1.6 | 0 | 0 | inf |   |   |  
test_add_dtype_float32_xnnpack_static_int8_per_channel | test_add_dtype
| xnnpack_static_int8_per_channel | {'dtype': torch.float32} | Pass |  
| TRUE | 0.66 | 0.818 | 7 | 0 | {'aten::add.Tensor': 1,
'quantized_decomposed::dequantize_per_tensor': 3,
'quantized_decomposed::quantize_per_tensor': 3} | {} | 2 | 0 | 0 | inf |
  |   |