Use _upb_Message_AlignedMemsetZero in upb_Message_Clear

`upb_Message_Clear` zeroed the message with a `memset` library call. Use the
same inline aligned zeroing as `_upb_Message_New` instead (message sizes are
multiples of kUpb_Message_Align, and messages are always
kUpb_Message_Align-aligned).

Also stop resetting the size of the message's internal data (unknown fields and
extensions) after clearing: the internal data pointer is in the message header,
which the clear zeroes, so nothing refers to that internal data afterwards and
the store was dead. (Copies never share internal data: upb_Message_ShallowCopy
gives the copy its own.) This saves loading the header before the clear.

android_arm64 (-c opt, fasttable), upb_Message_DeepCopy (which inlines
upb_Message_Clear), before:

  stp   x30, x23, [sp, #-0x30]!
  stp   x22, x21, [sp, #0x10]
  stp   x20, x19, [sp, #0x20]
  mov   x20, x2
  ldr   x8, [x0]
  ldrh  w2, [x2, #0x8]
  mov   x21, x1
  mov   w1, wzr
  mov   x19, x3
  mov   x22, x0
  and   x23, x8, #0xfffffffffffffffe
  bl    memset
  cbz   x23, 1f
  str   wzr, [x23]
1:mov   x0, x22
  mov   x1, x21
  mov   x2, x20
  mov   x3, x19
  bl    _upb_Message_Copy
  ldp   x20, x19, [sp, #0x20]
  cmp   x0, #0x0
  ldp   x22, x21, [sp, #0x10]
  cset  w0, ne
  ldp   x30, x23, [sp], #0x30
  ret

after (no call, so no callee-saved registers to spill around it, and no
load and store of the internal data):

  str   x30, [sp, #-0x10]!
  ldrh  w8, [x2, #0x8]
  mov   x9, x0
  add   x10, x0, x8
  str   xzr, [x9]
  cmp   x8, #0x8
  b.eq  2f
1:stp   xzr, xzr, [x9], #0x10
  stp   xzr, xzr, [x10, #-0x10]!
  cmp   x9, x10
  b.lo  1b
2:bl    _upb_Message_Copy
  cmp   x0, #0x0
  cset  w0, ne
  ldr   x30, [sp], #0x10
  ret

The error path of `_upb_Message_Copy`, which clears `dst`, likewise replaces
its `memset` call with the inline loop. No other function in the upb message,
wire and decode_fast objects changes.

PiperOrigin-RevId: 998754226