All the documents are loaded onto memory at once during UDFRunner.apply and this leads to a slowness if they exceeds machine's memory size

[HiromuHota]

Describe the bug

This is not a bug, but a memory-inefficient computing.
When executing UDFRunner.apply (like Parser.apply), all the documents are loaded onto memory at once (see the code below) before processing them in mutiple processes parallelly.

fonduer/src/fonduer/utils/udf.py

Lines 115 to 116 in 8134e4d

	for doc in doc_loader:
	in_queue.put(doc)

When len(doc_loader) is large, each file is large, and/or machine's RAM is small, this leads to memory paging and critically slows down the execution.

To Reproduce
Steps to reproduce the behavior:

1. Process a large set of documents on a machine with small RAM

Expected behavior

Load and process a document one at a time.
(When parallelism=N, load and process N documents at a time.)

Environment (please complete the following information):

• OS: N/A
• PostgreSQL Version: N/A
• Poppler Utils Version: N/A
• Fonduer Version: 0.8.2

Additional context
Add any other context about the problem here.