The extract() function passes tar entry names and symlink targets
through to the output stream without validating that they stay within
the extraction root. This allows a malicious container image to:

1. Write files outside the extraction directory via ../ prefixes
   (filepath.Clean preserves leading ../ sequences)
2. Create symlinks pointing to arbitrary host paths via absolute
   or ../-prefixed Linkname values
3. Use absolute paths in entry names to target fixed host locations

Add validation after filepath.Clean to skip entries with path
traversal (../ prefix) or absolute paths in both Name and Linkname
fields. This follows the mitigation pattern established by Docker
(moby/moby) and containerd after CVE-2018-15664 and CVE-2019-14271.

Add table-driven tests covering all five attack vectors (path
traversal, absolute path, symlink escape via absolute and relative
targets, hardlink escape) plus a positive case confirming safe
relative symlinks are preserved.

Absolute entry names (e.g. /etc/shadow) are now stripped of their
leading slash and emitted as relative paths rather than being silently
dropped.  This preserves existing behaviour for layers that store files
with absolute paths (used by crane edit fs and reflected in
TestEditFilesystem / TestCraneFilesystem) while still preventing
injection of host-absolute paths when consumers extract the tar to disk.

Symlink and hardlink entries with absolute or dot-dot targets continue
to be rejected outright, as those can never be safe.

Update TestExtractRejectsPathTraversal to document the normalization
behaviour, and update TestCraneFilesystem to use a relative path
consistent with what Extract now emits.

Fix golangci-lint staticcheck SA1019 warning.