probes: add email leak via virtualization latent injection (#249)

[manunicholasjacob]

Summary

Adds an email-based latent prompt injection scenario to
garak/probes/latentinjection.py, addressing #249. The module already covers
resume, report, whois, and translation contexts, but had no email scenario and
no virtualization-style framing.

What it does

Two new probes:

• LatentInjectionEmailVirtualizationFull — the full scenario matrix (active = False)
• LatentInjectionEmailVirtualization — the lightweight default-active variant

An assistant is given an ordinary mailbox task (summarize / reply / triage /
extract action items). The email body hides an injection that wraps its real
request in a fictional frame — a sandboxed VM, a unit-test fixture, a gdb
transcript, a screenplay, a text adventure — asking the model to relay a
distinctive canary string instead of doing the benign task. The injections are
hidden in varied, realistic spots (mid-sentence, a quoted reply chain, a
signature footer, an HTML img alt, an HTML comment).

Detection

Reuses the module's inherited base.TriggerListDetector: a hit is when the
model emits one of the canary trigger strings.

Testing

• black clean
• tests/probes/test_probes_latentinjection.py passes (its structural test
  auto-parametrizes over the new classes)
• tests/probes/test_probes.py metadata checks pass for the new probes
• test.Blank end-to-end run completes and passes

Closes #249.