PPC64: use pointer-width accesses when resetting string buffer

[mchoo7]

The PPC string lower/upper/reverse fast-function wrapper resets SBuf::w from SBuf::b using lwz/stw. Both fields are char pointers, so these 32-bit accesses are incorrect on PPC64.

This breaks the reset on big-endian PPC64, causing string.lower(), string.upper(), and string.reverse() to append to stale temporary buffer contents.

Use lp/stp so the accesses have the target pointer width.

[zhuizhuhaomeng]

Found one blocking issue relative to origin/v2.1-agentzh:

• [P1] Preserve the array search before calling tab_len_slow — src/lj_tab.c:672. The slow path assumes hi is a valid non-nil lower bound. Passing t->asize - 1 unconditionally violates that assumption when the array has trailing nil entries, returning allocated capacity instead of length. This breaks #t and operations such as table.remove. Retain the array-boundary search or adapt the slow path to handle trailing nil entries.

Confirmed in an isolated native build with the PPC64 C path enabled:

Check	Changed path	Original path
{1,2,3} after t[3] = nil	4	2
Empty table.new(100, 0)	100	0
Existing table_remove.lua	Fails	Passes

Can you check the review by AI @mchoo7

[mchoo7]

Found one blocking issue relative to origin/v2.1-agentzh:

• [P1] Preserve the array search before calling tab_len_slow — src/lj_tab.c:672. The slow path assumes hi is a valid non-nil lower bound. Passing t->asize - 1 unconditionally violates that assumption when the array has trailing nil entries, returning allocated capacity instead of length. This breaks #t and operations such as table.remove. Retain the array-boundary search or adapt the slow path to handle trailing nil entries.

Confirmed in an isolated native build with the PPC64 C path enabled:

Check Changed path Original path
{1,2,3} after t[3] = nil 4 2
Empty table.new(100, 0) 100 0
Existing table_remove.lua Fails Passes
Can you check the review by AI @mchoo7

This PR does not touch src/lj_tab.c. All it does is fixing wrong load/store size. Thus this PR has no relation with the AI review mentioned above.