Keep all turns and fix system prompt in ShareGPT conversations parsing

[SuperMarioYL]

What

Fix ShareGPT conversations handling in TextDataset._prepare_messages
(angelslim/data/text_dataset.py). The branch had two correctness bugs:

1. KeyError: 'system_prompt' — it guarded on data["system"] but then
   read data["system_prompt"]:

   if "system" in data and data["system"]:
       messages = [{"role": "system", "content": data["system_prompt"]}] + messages

   Any ShareGPT record that carries a system field crashes during data
   loading. (The sibling messages and input/output branches correctly
   guard and read system_prompt; only this branch was inconsistent.)

2. Multi-turn conversations were truncated to the first two turns and roles
   were assigned positionally:

   share_gpt_data = data["conversations"]
   messages = [
       {"role": "user", "content": share_gpt_data[0]["value"]},
       {"role": "assistant", "content": share_gpt_data[1]["value"]},
   ]

   A standard multi-turn ShareGPT record ([{"from": "human", ...}, {"from": "gpt", ...}, {"from": "human", ...}, ...]) lost every turn after the second,
   and the actual from of each turn was ignored. This defeats the downstream
   loop, which already scans for the last assistant turn and treats everything
   before it as prompt context — i.e. the pipeline is built for multi-turn data.

How

• Iterate over all turns, carrying each turn's role/content through the
  existing role-normalization loop (which already maps from→role,
  value→content, and human→user / gpt→assistant).
• Prepend the optional dataset-level system prompt (data["system"], with
  data["system_prompt"] accepted as a fallback), deduplicated against a
  conversation that already begins with a system turn so it is not doubled.
• Skip turns that carry no text payload, and harden the normalization loop so a
  turn missing a role no longer raises. The messages and input/output
  branches are left behavior-identical (the .get is a no-op for them).

Tests

tests/test_text_dataset_messages.py — CPU-only, no GPU / model weights / heavy
stack (the torch/transformers/datasets/pyarrow imports are stubbed;
_prepare_messages is pure-dict). Covers: system-field no-crash regression,
all-turns-preserved, leading-system-turn dedup, malformed-turn skip, and the
unchanged messages / input-output branches. The four conversations cases
fail on main and pass here; the two non-conversations cases pass on both.

Note on reuse

angelslim/compressor/speculative/train/data/data_utils.py already has a
convert_sharegpt_data helper with the same human→user / gpt→assistant
mapping. It is intentionally not reused here because it lives in the
speculative.train.data subpackage and pulls transformers.image_utils and
training-only deps at import — importing it into the core data loader would add
a heavier cross-package dependency than the few lines here. It also indexes
role_mapping[message["from"]] directly, which raises on a system turn,
whereas the normalization in text_dataset.py handles system turns.

Reviewer check: does keeping all turns + prepending the system prompt match
the intended ShareGPT calibration format, or should the loader continue to
support only single-turn conversations? The downstream _load_data loop
(scanning for the last assistant turn) suggests multi-turn is intended.