don't swallow stack traces

[jeremylightsmith]

this is a breaking change, but it will allow exceptions to retain their original stack traces when a retry fails

[snackycracky]

There's an inconsistency: only the exception path is changed to {e, stacktrace}, while the retriable-value path ({:retriable, e}) still passes the bare value. So an else clause now has to handle two different input shapes depending on whether the failure was a thrown exception or a returned value. The default clause only handles that by luck of the is_exception/1 guard falling through.

We can keep the stacktrace internal and reraise it only where the bug actually lives — the default else. User-supplied else clauses keep receiving the bare e, exactly as before.

exception_clause =
  if Keyword.has_key?(clauses, :else) do
    # user-supplied else: keep the historical contract (bare exception)
    quote do
      {:exception, e, _stacktrace} ->
        case e do
          unquote(else_clause)
        end
    end
  else
    # default: reraise with the original stacktrace instead of `raise e`
    quote do
      {:exception, e, stacktrace} ->
        reraise e, stacktrace
    end
  end

plus the one-line internal change in block_runner/2:

-  {:cont, {:exception, e}}
+  {:cont, {:exception, e, __STACKTRACE__}}

The {:exception, e, stacktrace} 3-tuple is purely internal and never reaches user code.

Why it's better

	this PR	non-breaking version
Original stacktrace preserved	✅	✅
Breaking change / major bump	❌ required	✅ avoided
Existing else tests need editing	❌ yes	✅ untouched
Heterogeneous else input	❌ yes	✅ no

The stacktrace test still passes unchanged:

test "reraises with the original stacktrace when giving up on an exception" do
  formatted =
    try do
      retry with: [1] do
        boom()
      end
    rescue
      e -> Exception.format(:error, e, __STACKTRACE__)
    end

  assert formatted =~ "(RuntimeError) boom"
  assert formatted =~ "RetryTest.boom"   # points at the failing site, not retry.ex
end

The one thing this PR gives that the non-breaking version doesn't: exposing the stacktrace to custom else clauses. If that's a goal, I'd suggest doing it additively (an opt-in second-element form) in a follow-up, rather than changing the shape everyone already depends on.