[FEA] Fix Parquet reader when using device_read_async while the destination device buffers are not ready

[ttnghia]

Similar to the issue with ORC reader with is fixed by #17074, Parquet reader also has similar problem with using device_read_async, when using device_read_async while the destination device buffers are not ready:

cudf/cpp/src/io/parquet/reader_impl_preprocess.cu

Lines 253 to 256 in e41dea9

	auto buffer =
	rmm::device_buffer(cudf::util::round_up_safe(io_size, BUFFER_PADDING_MULTIPLE), stream);
	auto fut_read_size = source->device_read_async(
	io_offset, io_size, static_cast<uint8_t*>(buffer.data()), stream);

We need to fix it ASAP otherwise there is a risk of data corruption.

From the implementation of datasources, I see that this bug only affects reading data from files. In such situations, the host-2-device memcpy ops do not use user-provided stream. Copying data from other host/device buffers to device buffer is not affected.