Fix some mania replays failing to import

[bdach]

Found during testing of #9988.

Summary

At the start of each replay, stable adds two special replay frames: one at time 0, and one at the boundary of the skip to gameplay (minus one millisecond). The first frame was handled already:

osu/osu.Game/Scoring/Legacy/LegacyScoreDecoder.cs

Lines 247 to 250 in 93a9994

	if (i == 0 && diff == 0)
	// osu-stable adds a zero-time frame before potentially valid negative user frames.
	// we need to ignore this.
	continue;

but the second wasn't. That same frame also happens to specify a position of (256, -500), which the mania decoder interpreted as belonging to the 9th column (as 256 == 1 << 8, and the shifts are zero-based).

Not all maps were affected by this, however; if a map doesn't have a skip at the start, the second frame has a time of -1, which would end up being caught by the second guard right underneath the first. In short, all replays that:

• were for a mania map,
• had below 9 keys,
• whose associated map had a skip forward at the start,

would fail to import, catching this exception.

Resolve by replacing the existing check by a more directly targeted one.

Remarks

A rudimentary smoke test case is attached, to demonstrate that importing no longer crashes. I manually created this replay on stable, using a simple custom 7K beatmap. I actually found another bug during writing the test case, which will be addressed in a PR coming up ahead.

The check is quite ugly and brittle, but I'm not seeing another way of scoping this better. It shouldn't swallow frames it shouldn't, I think (stable replays are protected by the i < 2 index check, and lazer replays are protected by the Y check - there's nothing that has the right to return a meaningful frame with negative Y).

I'm not even sure if lazer has the same notion of a skip boundary that stable does, and even if it does, I'm not sure relying on that is any better than this.