Prepare the repository for public visibility

[deepfates]

A stranger who opens this repository now finds an install line they can run,
a stated license position, a pull request template that asks for checks they
can actually run, and a bug template. Nothing tells them the repository is
private or that they need credentials, because neither is true.

Each change:

• README's Install section installs from the public repository at a tag, with
  branch: "main" named as the way to get Imp.ACP and Imp.MCP.connect/2
  until a tag containing them is cut; the "Release state" paragraph at the top
  now says the same thing.
• README's Install section adds one sentence for OTP release users: ex_mcp is
  runtime: false, so a release using the adapters needs
  applications: [ex_mcp: :load], linked to the existing wording in
  docs/PRODUCTION_OPERATIONS.md rather than duplicated.
• RELEASE_NOTES.md and examples/deployment/README.md drop the same private
  repository and credential statements; the five Livebooks' standalone install
  comments drop the word "private" too.
• decisions.md: the 2026-08-23 Hex ruling's status column records that the
  repository is public as of 2026-09-17 and that the Hex question is open.
• New NOTICE names DSPy, its MIT license and copyright holder, and the two
  files ported from it (lib/imp/optimizer/utils.ex, lib/imp/adapter/chat.ex).
  README points at it; mix.exs package_files/0 ships it.
• .github/pull_request_template.md now asks an outside contributor for
  mix check and a plain statement of what changed and how it was verified,
  and says provider-backed checks are the maintainer's to run.
• Every third-party action in ci.yml, evidence.yml and the elixir-setup
  composite action is pinned to a full commit SHA with its version tag in a
  trailing comment, at the versions Dependabot Bump actions/cache from 4 to 6 #89, Bump dorny/paths-filter from 3 to 4 #90 and Bump actions/setup-python from 5 to 7 #91 propose
  (actions/cache v6.1.0, dorny/paths-filter v4.0.3, actions/setup-python
  v7.0.0), plus actions/checkout v7.0.1, erlef/setup-beam v1.24.1 and
  denoland/setup-deno v2.0.5. Those three Dependabot pull requests can be
  closed. The ci.yml header comment says the pins move with their comments.
• New .github/CODEOWNERS (* @deepfates) and
  .github/ISSUE_TEMPLATE/bug.md. No feature template.
• AGENTS.md keeps its design orientation but stands alone: no reference to a
  workshop file or to sibling repositories, and it points at CONTRIBUTING.md
  for the gates and decisions.md for the rulings. I kept it rather than
  folding it into CONTRIBUTING.md because CONTRIBUTING.md covers setup and
  gates while AGENTS.md is design context, and merging them would bury both.

What I could not do: neither open Dependabot alert can be fixed. The nltk
alert (high, benchmarks/requirements-ifbench-parity.txt) and the accelerate
alert (medium, priv/trl_worker/uv.lock) both report
first_patched_version: null — no fixed release exists, and both pins are at
the top of the vulnerable range. Separately, the nltk pin is a source-exact
IFBench scorer environment that CI builds to reproduce upstream scores, so
even a patched release would have to be taken deliberately as a parity change,
not as a dependency bump.

One test assertion changed: test/package_contract_test.exs had a test named
"README gives the usable private source install"; the name is now "the usable
source install". Its assertions were already true and are unchanged. NOTICE
was added to that file's repository and product file lists so packaging is
asserted.

Verification, from this branch:

mix format --check-formatted          clean
MIX_ENV=test mix compile --warnings-as-errors   Generated imp app
mix docs                              generated, no warnings
mix test test/documentation_contract_test.exs test/package_contract_test.exs
                                      41 tests, 0 failures
ruby -ryaml -e YAML.load_file         parses: ci.yml, evidence.yml,
                                      .github/actions/elixir-setup/action.yml

Files owned by the two sibling pull requests (the evidence-bookkeeping
deletion and the benchmark surface) are untouched.