[roll] Roll fuchsia [memory][scudo] Raise CustomSizeClassMap max size to 128 KiB

With MaxSizeLog = 16 (MaxSize = 65,552 B), allocations slightly above
64 KiB (including the 16-byte Scudo chunk header) overflow the Primary
allocator into the uncached Secondary allocator (MapAllocatorNoCache),
causing a per-call VMO create/map/unmap/close cycle and zero-page
faults:
- fdf_arena::Allocate(65536) (65,568 B needed) for driver FIDL tables
  and unbounded vector requests
- zxio_dirent_iterator_t (65,616 B needed) on directory iteration
- Sdmmc.Request client sync buffer (65,600 B needed)
- HLCPP IncomingMessageBuffer (66,576 B) and OutgoingMessageBuffer
  (67,088 B)

Raise MaxSizeLog to 17 (128 KiB), drop the 3248-byte class to keep the
total class count at 44 (NumClasses = 45), and replace the 65552-byte
tail class with 67600 (65536 + 2*1024 + 16, satisfying the
TableSizeClassMap step invariant for (64 KiB, 128 KiB]) and 131088
(128 KiB + 16 B). This retains 99.2% of the heap padding savings while
serving 64 KiB+ header-padded and 128 KiB buffers from Primary.

This was discovered after Icaef2bae5b806154d82b92a6615e7a468837f9ff
was implicated in several storage-related performance regressions
(see bugs below).

Original-Original-Bug: 382623426
Original-Original-Bug: 579983314
Original-Original-Bug: 579983312
Original-Original-Bug: 579974039
Original-Original-Bug: 579976544
Test: fx test scudo-unittests
Original-Original-Reviewed-on: https://fuchsia-review.googlesource.com/c/fuchsia/+/1944448
Agent-Code-Review: Fuchsia CL Risk (Advisory) <groupagent-fuchsia-risk@google.com>
Risk: Fuchsia CL Risk (Advisory) <groupagent-fuchsia-risk@google.com>
Original-Original-Revision: 1a9d3a51afcc8904befa58f61493a7b89059cbba
GitOrigin-RevId: 07a650052a43b40ce7ccd2d2b956e0d6b93e4f77
Change-Id: I74aae8ae8c563cc106d9c7a3434bb44434f587a6
Reviewed-on: https://fuchsia-review.googlesource.com/c/integration/+/1955297
Cr-Commit-Position: refs/heads/main@{#211449}