ext4: scale default block size by filesystem size

[eriknordmark]

Summary

recalculateBlocksize was a placeholder: every bracket of its size-based
switch set the same sectorsPerBlock = 2, so a default-parameter
Create always used 1 KiB blocks regardless of filesystem size. On
filesystems of a few GiB the journal allocator — capped at 128 MiB =
131072 1-KiB blocks — asks for more blocks than the per-extent cap
(maxBlocksPerExtent = 32768) or the inode's 4-extent root limit can
fit, and Create fails with

could not initialize journal: cannot allocate more than 65535 blocks in
a single extent

(see #402 for the trace).

This change makes recalculateBlocksize actually pick a block size by
filesystem size, matching mke2fs's small-vs-default split: 1 KiB
blocks below 512 MiB, 4 KiB blocks at or above. With 4 KiB blocks a
128 MiB journal needs exactly maxBlocksPerExtent (32768) blocks and
fits in a single extent.

Because resize_inode / reserved-GDT growth in go-diskfs currently
only supports 1 KiB blocks (there's an existing TODO: Properly support resize_inode at the GDT-sizing block, and
TestCreateWithBlockSizes already passes
WithFeatureReservedGDTBlocksForExpansion(false) for non-1 KiB cases),
this also drops the feature when we picked the larger default. A
caller who explicitly sets SectorsPerBlock is unaffected: they pass
their own block size and the default-only path doesn't run.

Test

Added TestCreateMultiGB — Create on 1 GiB and 2 GiB images, validated
with e2fsck -f -n. Both pass with the fix; the 2 GiB case is the
size that reproduces #402 on master. Full go test ./... is green.

Behavior change to flag

This is a visible default change: existing callers who got 1 KiB
blocks back from Create on a >512 MiB filesystem will now get 4 KiB
blocks. Almost certainly the right default (matches mke2fs), but
worth calling out. Callers who need the old behavior can set
Params.SectorsPerBlock = 2 explicitly.

Not in this PR

Large-file writes that need >32768 blocks still hit the same inode
4-extent root limit fixed by no other change — createRootExtentTree
explicitly returns "cannot create root internal node" past 4 entries.
That's a separate, pre-existing limitation; this PR closes #402 for
the journal-allocation case without addressing it. Worth filing as its
own issue.

Fixes #402.