Improve performance of `os.walk()`

[barneygale]

There are a couple of minor performance improvements possible in os.walk():

• We don't need to manually pump the os.scandir iterator, given we handle exceptions from next() like exceptions from scandir() itself, i.e. by ignoring the problematic directory and moving on. We can use a for loop like filthy casuals.
• In bottom-up mode, we can handle exceptions from entry.is_symlink() in the same block as those from entry.is_dir(), which avoids a few temporary variables.
• In top-down mode, we can call os.path.join() once on a parent directory rather than for each child path.

Linked PRs

• GH-119169: Speed up os.walk(topdown=False) #119186
• GH-119169: Skip reversing sibling directories in os.[f]walk(topdown=False) #119473
• GH-119169: Implement pathlib.Path.walk() using os.walk() #119573
• [3.13] GH-119169: Implement pathlib.Path.walk() using os.walk() (GH-119573) #119750
• GH-119169: Slightly speed up os.walk(topdown=True) #121431
• GH-119169: Simplify os.fwalk() exception handling #121432
• GH-119169: Speed up os.fwalk(topdown=False) #121433
• GH-119169: Simplify os.walk() exception handling #121435

[barneygale]

In bottom-up mode, we can handle exceptions from entry.is_symlink() in the same block as those from entry.is_dir(), which avoids a few temporary variables.

Upon closer inspection, I think I was wrong. Symlinks to directories where is_symlink() raises should still be categorised as dirs.