simd: stop the AVX2 rolling checksum reading past its buffer

The loop is software-pipelined: each iteration folds in the 64 bytes it
preloaded last time and preloads the next 64.  Nothing stopped the final
iteration doing that preload, so it always read the 64 bytes after the region
it was asked to checksum.

Not an edge case.  The assembly processes len&~63 and leaves the remainder to
the caller, so the remainder is by construction under 64 bytes and the preload
passed buf+len on every call, by 64 minus the remainder.

It normally landed in slack inside the map_ptr() window and nothing noticed.
Where the buffer ended near an unmapped page it was a SIGSEGV in the middle of
a transfer -- reported on macOS x86-64 by Roland Kletzing, whose `partial` run
died with "connection unexpectedly closed" because the generator had crashed.
A guard page reproduces it on Linux too, so it was latent there, not absent.

Run the pipelined loop one block short and finish the last block in .last,
which does the same arithmetic without the preload.  No per-iteration cost, and
checksums are bit-identical -- simdtest compares every implementation against
the C reference.

The earlier fix for that report, "lib: use .balign in md5 x86-64 asm", was to
the md5 assembly.  It addressed the linker alignment warning that appeared
alongside, not this.

simd-checksum-avx2.S

@@ -46,6 +46,14 @@ get_checksum1_avx2_asm:
 	vpxor	xmm1, xmm1, xmm1 # reset both partial sums accumulators.
 	vpxor	xmm4, xmm4, xmm4
 	mov	eax, [r8]
+	# The loop below is software-pipelined: it preloads the 64 bytes AFTER the
+	# ones it is folding in.  Run it one block short and finish that block in
+	# .last, which does not preload -- otherwise the final iteration reads 64
+	# bytes past the end of what the caller gave us.  The remainder (len & 63)
+	# is always under 64, so that read always crossed buf+len; it only faulted
+	# when the buffer happened to end at a page boundary.
+	sub	esi, 1
+	jz	.last
 	.p2align 4 		# should fit into the LSD allocation queue.
 .loop:
 	vpmaddubsw	ymm0, ymm15, ymm2  # s1 partial sums
@@ -70,11 +78,33 @@ get_checksum1_avx2_asm:
 	vpaddd ymm6, ymm10, ymm6  #  32*CHAR_OFFSET
 	vpaddd ymm1, ymm13, ymm1  # 528*CHAR_OFFSET
 #endif
-	vmovdqa ymm2, ymm8   # move the next 64 bytes 
+	vmovdqa ymm2, ymm8   # move the next 64 bytes
 	vmovdqa ymm3, ymm9   # into the right registers
 	sub	esi, 1
 	jnz	.loop
 
+	# The final 64 bytes, already in ymm2/ymm3.  Same arithmetic as the loop
+	# body with the preload, the prefetch and the loop control dropped.
+.last:
+	vpmaddubsw	ymm0, ymm15, ymm2  # s1 partial sums
+	vpmaddubsw	ymm5, ymm15, ymm3
+	vpaddd	ymm4, ymm4, ymm6
+	vpaddw	ymm5, ymm5, ymm0
+	vpsrld	ymm0, ymm5, 16
+	vpaddw	ymm5, ymm0, ymm5
+	vpaddd	ymm6, ymm5, ymm6
+	vpmaddubsw	ymm2, ymm7, ymm2  # s2 partial sums
+	vpmaddubsw	ymm3, ymm12, ymm3
+	vpaddw	ymm3, ymm2, ymm3
+	vpsrldq	ymm2, ymm3, 2
+	vpaddd	ymm3, ymm2, ymm3
+	vpaddd	ymm1, ymm1, ymm3
+
+#if CHAR_OFFSET != 0
+	vpaddd ymm6, ymm10, ymm6  #  32*CHAR_OFFSET
+	vpaddd ymm1, ymm13, ymm1  # 528*CHAR_OFFSET
+#endif
+
 	# now we reduce the partial sums.
 	vpslld	ymm3, ymm4, 6
 	vpsrldq	ymm2, ymm6, 4