buffs: add FormatSpread prompt-format-sensitivity buff (#396)

[manunicholasjacob]

Summary

Adds a formatspread buff (garak/buffs/formatspread.py) motivated by
FormatSpread (Sclar et al., ICLR 2024): models are sensitive to spurious prompt
formatting. Addresses #396.

What it does

FormatSpread rewrites each prompt into a fixed, deterministic set of surface-
format variants composed from the paper's format grammar: descriptor
(Instruction / Task / Request), separator (": ", "::", " - ", "\n"),
casing (identity / upper / lower / title), and field spacing ("\n\n", " || ")
with an optional answer descriptor. Buffing a probe with it surfaces how much
formatting alone moves model behavior.

Detection

None added. Buffs transform prompts; detection is handled by the probe or
detector the buff is applied to.

Testing

• black clean, pylint 10/10
• tests/buffs/test_buffs.py and test_buff_config.py pass (structural and
  load/transform tests auto-cover the new buff)
• end-to-end: buffing a real probe (latentinjection.LatentWhois) expands
  prompts 12x and completes

Closes #396.