[CameronBrooks11]

This PR adds blockspec.edit, so a program can change one block of a document without rewriting its lines itself (closes #69). The design is in docs/editing-api.md.

• replace_block(text, node, source, /), insert_block(text, source, /, *, before=None, after=None) and delete_block(text, node, /). Each is a pure function from text to text. With neither before= nor after=, an insert goes after the last block.
• Naming a block. A block is named by its node path [i], the path parse reports. A caller holding a name resolves it with parse(text).resolve(name)["node"].
• Lists, block quotes and tables. A path into a container is refused as nested-node. A whole list, quote or table can be replaced, inserted or deleted as one block, but no item or cell can. Writing inside a container means writing its prefixes, and §11.7's write-back shows how many of those lines draft 0 cannot write.
• Read-back. The input and the result are both read with the reader, so §9 refusals come through with their own codes. The API's own codes are bad-position, no-such-block, nested-node, not-one-block, changes-another-block and rebinds-name, raised as EditRefused, a Refusal.
• Unchanged text. A line the operation is not about keeps its bytes and its line ending. <doc>.notes/ records are never touched.

Cases. The cases are in tests/edit-cases/, not conformance/cases/, because §2 leaves editor behaviour unspecified. An author who worked from the design and SPEC alone, without the reference implementation, wrote 113 cases: 18 replace, 29 insert, 20 delete and 46 refusal, each refusal with a control. All passed on the first run with no case changed. docs/editing-api.md was clarified after the cases were written. tests/test_edit.py adds 9 tests, including an insert into an empty file, which ends its line with LF. just check exits 0, and just test passes 305.

Two checks no case reaches. I mutated each refusal and checked that the file's hash differed every time:

• Killed: bad-position, no-such-block, nested-node, both not-one-block checks, rebinds-name, and changes-another-block with its count and frontmatter checks removed together.
• Survived: two backstops inside changes-another-block. One is the frontmatter clause alone, the other the per-block comparison. The block-count check catches every case that reaches them, and I found no input that reaches either one alone. I kept them because they guard against an edit silently changing another block.

Inserting inside a region. A block inserted between a region's markers joins that region, which is what the markers mean. rebinds-name refuses an edit that would make a name bind to something else. For example, inserting a block directly after a one-line marker would move the marker's name onto the new block.

Signed-off-by: Cameron Brooks 95648474+CameronBrooks11@users.noreply.github.com