perf(stdlib): defer string.format width padding to iolist

[davydog187]

Context

string.format with width-flagged specifiers is 2.03× slower than luerl
(width-flagged, n=1000: 5.51ms vs 2.71ms in benchmarks/string_format.exs).
The literal-heavy fast path from PR #299 is now winning by 2.6×, so the
remaining work is in per-specifier width padding.

apply_width_flags/3 at lib/lua/vm/stdlib/string.ex:828 builds
String.duplicate(pad_char, deficit) and concatenates with <>, producing a
fresh binary per specifier. Luerl returns [Padding, Bin] as iolist
(deps/luerl/src/luerl_lib_string_format.erl:303-319) and lets the final
iolist_to_binary materialise everything at once.

Change

lib/lua/vm/stdlib/string.ex:828-859 plus its caller in format_string/3.

• apply_width_flags returns iolist ([pad, str] / [str, pad]) instead
  of a binary.
• Thread iolist through format_string so the existing accumulator from
  PR perf(stdlib): iolist string.format and plain-table sort/concat fast paths #299 absorbs the padding.
• Materialise once at the top level via IO.iodata_to_binary/1.

Verification

• mix test test/lua/vm/stdlib/string_test.exs
• Full mix test
• mix run benchmarks/string_format.exs

Expected

Closes most of the 1.5–2× width-flag gap.