Base64 data URIs fail or decode incorrect bytes when their payload is percent-encoded

Bug

On current main cc0ca9e, percent-encoded characters in a base64 data URI reach the base64 decoder without URL decoding. Some valid inputs raise binascii.Error; others silently produce incorrect bytes.

No network or fixture is needed (Python 3.12):

from markitdown._uri_utils import parse_data_uri

print(parse_data_uri("data:text/plain;base64,SGVsbG8%3D"))
# Expected: ('text/plain', {}, b'Hello')
# Actual: binascii.Error

print(parse_data_uri("data:application/octet-stream;base64,%2B/8="))
# Expected bytes: b'\xfb\xff'
# Actual bytes: b'\xd8\x1f\xfc'

MarkItDown.convert_uri also fails on URL-escaped base64 payloads. Raw base64 and non-base64 data URIs already work and should keep their behavior. RFC 2397 section 3 defines the data field using URL characters; decoding those escapes must precede base64 decoding. The new regression suite reports 16 failures and 2 passes against unchanged production code.

Start with parse_data_uri in markitdown._uri_utils and trace how MarkItDown.convert_uri handles data URIs. Run the regression cases described in the issue, then verify that percent-encoded base64 payloads decode to the expected bytes while raw base64 and non-base64 data URIs retain their existing behavior.