Start QOI decoding with a zero-initialized array of previously seen pixels

[radarhere]

This originated in discussion at #9007

When starting to decode a QOI image, https://qoiformat.org/qoi-specification.pdf states that

The decoder and encoder start with {r: 0, g: 0, b: 0, a: 255} as the previous pixel value.

Pillow/src/PIL/QoiImagePlugin.py

Lines 43 to 54 in b844007

	def _add_to_previous_pixels(self, value: bytes | bytearray) -> None:
	self._previous_pixel = value
	r, g, b, a = value
	hash_value = (r * 3 + g * 5 + b * 7 + a * 11) % 64
	self._previously_seen_pixels[hash_value] = value
	def decode(self, buffer: bytes | Image.SupportsArrayInterface) -> tuple[int, int]:
	assert self.fd is not None
	self._previously_seen_pixels = {}
	self._add_to_previous_pixels(bytearray((0, 0, 0, 255)))

However, the part of the specification describing _previously_seen_pixels states that

A running array[64] (zero-initialized) of previously seen pixel values is maintained by the encoder and decoder.

While I don't think these two statements from specification fit together in the most natural way, the code that currently exists is not 'initialized' at zero exactly - we're starting out with the previous pixel value of (0, 0, 0, 255). This PR stops that.