[Conv] Fix Conv L1 Estimation for Automatic DRAM Slicing

[wransom-TT]

During implementation of automatic DRAM slicing for Pool2D it was discovered that Conv2D sometimes runs with an invalid slice config but does not crash with OOM errors, indicating that the Conv2D L1 estimation is incorrect.

To repro run:
pytest "tests/ttnn/unit_tests/operations/conv/test_conv_transpose2d.py::test_convt2d_dram[mirror_kernel=False-preprocess_weights=False-activations_dtype=DataType.BFLOAT8_B-layout=Layout.TILE-weights_dtype=DataType.BFLOAT16-batch_size=1-input_height=512-input_width=512-input_channels=64-output_channels=64-filter_height=3-filter_width=3-stride_h=1-stride_w=1-pad_h=1-pad_w=1-out_pad_h=0-out_pad_w=0-config=None-shard_layout=TensorMemoryLayout.HEIGHT_SHARDED-num_slices=4-slice_type=SliceTypeEnum.DRAMSliceWidth-device_params={'l1_small_size': 65536}]"

on Wormhole N150 after removing the conv_bypass.