codegen fixes, --no-gc-check / --fast-math, raise caps, DATE$/TIME$ shift

Four roadmap items:

- codegen: fix parenthesized string comparison.  emit_atom didn't
  consume the body of a string-literal token (`"`), so for
  PRINT (A$+B$ < "ZZZ") it emitted a 0 placeholder, advanced one byte,
  and left "ZZZ" to be reparsed as a variable + extra trailing tokens
  -- the binary then failed to link with `var_ZZ_sng` undeclared.
  emit_atom now skips to the closing quote.  Separately, the
  left_type tracking in emit_num_prec dropped VT_STR after a string +
  string concat (becoming VT_SNG), so the string-comparison codepath
  skipped when the relational operator arrived.  Preserve VT_STR
  through TOK_PLUS when both operands are strings.  Verified: paren
  string-cmp now compiles and produces the same -1 / 0 result as the
  interpreter.

- compiler: --no-gc-check and --fast-math optimization flags.
  --no-gc-check skips the per-line gwrt_check_line() (no string-pool
  GC, no Ctrl+Break trap).  --fast-math drops the divide-by-zero
  guard on `/`; the divisor still goes through (double) so 10/0
  produces inf rather than SIGFPE.  Both threaded through
  codegen_opts_t and exposed in --help.  --inline-arrays from the
  roadmap deferred -- larger refactor.

- interp: raise static caps on 32-bit / Linux builds.  vars 256
  -> 1024, arrays 64 -> 256, MAX_FOR_DEPTH 16 -> 64, MAX_GOSUB_DEPTH
  24 -> 128, MAX_WHILE_DEPTH 16 -> 64.  Codegen FOR_STACK_MAX 16
  -> 64.  Analysis-pass caps: MAX_LINES 4096 -> 8192, MAX_VARS 256
  -> 1024, MAX_GOTOS 256 -> 1024, MAX_DATA 1024 -> 4096,
  MAX_GOSUB_RET 256 -> 1024.  16-bit DOS keeps the original modest
  caps via #ifdef _M_I86 -- the MEDIUM model has a single 64KB
  DGROUP for all static data and the bumped sizes broke runtime
  startup under DOSBox-X.  16-bit binary grew from 128KB to 132KB
  from the offset_secs field plus DATE$/TIME$ shift code, well
  within the FreeDOS budget.

- interp + codegen: DATE$ / TIME$ assignment via process-local
  clock offset.  Was a no-op accept-and-ignore.  Now sets
  gw.time_offset_secs (long), and DATE$ / TIME$ / TIMER readers
  apply it to time(NULL) before formatting.  The OS clock is
  unaffected (would need root).  Compiled-binary readers also
  reference gw.time_offset_secs since libgwrt shares the gw
  struct.  Verified: PRINT DATE$; DATE$="12-31-1999"; PRINT DATE$
  shows the expected before/after in both interpreter and AOT
  paths.

After these changes: 72/72 interpreter tests, 68/68 compat, 63/63
compiler tests, DOS smoke under DOSBox-X all pass.  Build clean on
both Linux (cmake) and 16-bit DOS (build_dos.sh 16).

docs/roadmap.md

@@ -112,8 +112,12 @@ Tested on FreeDOS 1.4 via QEMU.
 
 ## Known Limitations
 
-- Maximum 256 variables, 64 arrays, 16 FOR nesting, 24 GOSUB nesting,
-  16 WHILE nesting
+- Static caps -- 32-bit / Linux builds: 1024 variables, 256 arrays,
+  64 FOR nesting, 128 GOSUB nesting, 64 WHILE nesting.  16-bit real-mode
+  DOS keeps the original modest caps (256 / 64 / 16 / 24 / 16) because
+  the MEDIUM model has a single 64KB DGROUP for all static data.
 - `CALL`/`CALLS` (machine code execution) raises Illegal function call
-- `DATE$`/`TIME$` assignment accepted but does not modify the system clock
+- `DATE$`/`TIME$` assignment shifts the program's view of the clock via
+  a process-local offset; the OS time is unaffected (setting the OS
+  clock would require root)
 - Device stubs (`ERDEV`, `IOCTL`, `COM`, `LCOPY`) return defaults