Today `ilab generate` creates a large percentage (> 50% in my testing)
of discarded examples because the examples do not match our expected
format. Typically, the model fails to follow our complicated numbering
scheme for instructions, inputs, and outputs. This change simplifies
that scheme, taking inspiration from Emacs org-mode without any rigid
adherence to it.

Before this change, we expected the model to continue a format like:

```
1. Instruction: Generate a joke involving three horses.
1. Input:
<noinput>
1. Output:
There once were three horses...
```

After this change, we now give the model:

```
* Task 1
** Instruction
Generate a joke involving three horses.
** Input
<noinput>
** Output
There once were three horses...
```

In my local testing, the model is able to replicate our new format
with a much higher accuracy than the previous format, resulting in
substantially fewer discarded generations due to format.

Other formats I considered but decided against:
- JSON was excluded because the model was not able to reliably handle
quotes or special characters when generating JSON, even though overall
it did seem to understand the expected format.
- XML was excluded because it resulted in a more verbose prompt,
consuming quite a bit more of our available context window. The model
did seem to have a good grasp of XML with CDATA elements used to
handle special characters.
- Markdown was excluded because knowledge documents are in Markdown
format, and I wanted to minimize the chances of the embedded knowledge
document interfering with the expected parsing structure.

Signed-off-by: Ben Browning <bbrownin@redhat.com>

Occasionally the model inserts a trailing colon in the generated
example headings, so optionally allow that in the regular expression
that splits based on those headings. This is another incremental
decrease in the overall number of discarded instructions with this new
format.

Signed-off-by: Ben Browning <bbrownin@redhat.com>

I missed this in the prior commits, but this updates the stop sequence
passed to the Completions API to match our new task format so that the
model only generates two new instructions per run. We ask it to
generate five, but seed it with two examples, it generates two, and it
stops as soon as it starts generating the fifth.

Signed-off-by: Ben Browning <bbrownin@redhat.com>