[khromenokroman]

@olofhagsand Hello

Description:
Fixes #698

Problem

struct xp_ctx (xc_invariant) and struct xpath_tree (xs_cached_result,
xs_cached_initial) defined some fields only under #ifdef XPATH_CURRENT_OPTIMIZE.
That macro comes from clixon_custom.h via clixon_config.h, which is not
included by:

• lib/src/clixon_xpath_ctx.c
• external code (plugins) using the installed headers via <clixon/clixon.h>

So the struct layout differed between compilation units. ctx_dup() in
clixon_xpath_ctx.c allocated xp_ctx without xc_invariant, while
clixon_xpath_eval.c read and wrote that field past the end of the allocation.

On x86_64 this is hidden by struct padding (both variants are 72 bytes). On
32-bit ARM the sizes are 48 vs 56 bytes, so the heap gets corrupted, which
explains the random wrong NACM decisions reported in #698.

Reproduction

Build master with AddressSanitizer and __attribute__((packed)) on
struct xp_ctx (as suggested in #698) to remove the padding, then validate a
leafref whose path uses current(), e.g.
/group[name=current()/../group]/member/id:

  ERROR: AddressSanitizer: heap-buffer-overflow
  READ of size 4 ... 0 bytes after 64-byte region
      #0 xp_eval clixon_xpath_eval.c:1595
      ...
  allocated by thread T0 here:
      #1 ctx_dup clixon_xpath_ctx.c:97

Fix

1. Define the fields unconditionally, so that the struct layout is the same in
   all compilation units and in external code regardless of compile options.
   The cost when XPATH_CURRENT_OPTIMIZE is disabled is one int in xp_ctx
   and two pointers in xpath_tree; the fields simply stay zero.
2. Include clixon_config.h in clixon_xpath_ctx.c, like the other xpath
   sources do.

(1) is the actual fix; (2) alone would fix the library but leave plugins
compiled against the installed headers with a different layout.

Testing

• ASan + packed xp_ctx: overflow reproduced on master, gone with this PR.
• Builds and validates correctly with XPATH_CURRENT_OPTIMIZE both defined
  and undefined.
• Regular library build has no new warnings.

No regression test is added: on x86_64 the bug is masked by padding, so a
test would not detect it there.