0.9.0

A restore shares its snapshot's image instead of copying it, script requests
are faster by default, a worker can run compiled with one option, a loaded
module no longer keeps its whole input file in memory, and the CPython reactor
ships stripped with its standard library precompiled. Workers built on
py_reactor.wasm need the new build. Drop recycle from wasm:restore/3 calls
and recycle_idle from worker options; both are now ignored.

• Restored memory shares its image. A restore no longer copies the
  image: the memory reads the image's pages in place and copies a 4 KiB page
  on its first write. A restore costs a page table and a request the pages it
  writes: a CPython restore is 0.79 ms instead of 12.0, and a node holds 124
  CPython instances under page_limit 4096 instead of 6. Loading a snapshot
  file costs what it did.
• Restore recycling is removed. wasm:restore/3 no longer takes recycle,
  and the worker option recycle_idle is gone. Nothing replaces them: every
  restore now shares the image.
• Compiled code reaches restored memory through a translation cache. A
  compiled request's guest time is 12% to 49% higher than with a copied image;
  whether the whole request is faster depends on how much of it the restore
  was (a CPython request with an entry: 5.4 ms to 3.0 ms; a compiled QuickJS
  script request: 6.5 ms to 7.7 ms). Generated code is ABI 14, so the compiled
  tier's disk cache is rebuilt once.
• The node page budget counts allocated pages. A restored memory charges
  its page table and the pages it writes; any memory charges its growth in
  whole chunks, so a three-page memory counts four. max_memory_pages still
  bounds the size a guest sees.
• A write can now be refused for want of budget, with exhaustion /
  memory_limit, leaving every byte as it was. A restore that cannot be
  afforded is refused the same way; it used to surface as malformed /
  internal.
• The snapshot budget charges whole pages: 64 KiB for each page that holds
  data and 8 bytes per page of address space, held while the image or any
  memory restored from it remains.
• A snapshot file is checked against its module before it is built. One
  with runs out of order or overlapping, or memories outside the module's
  limits, is refused by name; one over the budget, or mapping more than 2^20
  pages in one memory, is refused as exhaustion / snapshot_budget.
• Default heap floors. A new optional adapter callback, defaults/1, is
  given the worker's resolved limits and answers runner_min_heap_words and
  capture_min_heap_words when the caller did not set them. wasm_lua and
  wasm_javascript default the runner floor to 200,000 words; wasm_python
  to 1,500,000 on the compiled tier (compile => true, fuel => infinity)
  and 1,000,000 otherwise. Pass the option to change it, or 0 to turn it
  off. A floor is held only while a request executes; docs/tuning.md has
  what it costs a pool and how +MMmcs affects RSS. No adapter sets a
  capture floor.
• Lowered code is shared per module. A module loaded through the module
  cache publishes its validation context, function table and the function
  bodies requests reach, once per node, so a request in a fresh runner no
  longer lowers them again. Nothing to set.
• New worker option compiled. compiled => true sets fuel => infinity, compile => true, compile_after => 1 and compile_quality => baseline under your own limits. With a finite fuel the start fails with
  {bad_option, compiled, #{fuel => N}}; that combination used to interpret
  without a word. Metered stays the default.
• wasm:restore/3 expands profile as wasm:instantiate/3 does. A
  restored instance under profile => script used to keep the default compile
  threshold of 32.
• A decoded module owns its bytes. Data segments, names and the custom
  sections it keeps are copied out of the input, and .debug_* sections are
  dropped, so the input binary is freed when the caller lets go of it. A
  module used to pin the whole file: 31 MB for the CPython reactor.
• Fixed: the per-process function cache ignored fuse, so an unfused
  instance could hand its bodies to a fused one in the same process.
• bench/paths/ gains gap.erl, guestprof.erl, lowbench.erl and
  retainbench.erl; reqbench's QuickJS and Lua arms now run at infinite fuel,
  and so reach the compiled tier.
• scripts/build-python-reactor.sh strips the DWARF with
  wasm-opt --strip-debug: 7.4 MB instead of 30.9, at the same request
  speed. The build now needs binaryen's wasm-opt, or WASM_OPT set to it.
• The standard library ships precompiled, as unchecked hash-based .pyc
  files compiled by the build's own host interpreter. A module a request
  imports that is not in the image loads instead of compiling from source.
  Editing a .py under the library has no effect until you rebuild it.
• The reactor freezes its image out of the cyclic collector. init(), and
  an entry's capture, end with gc.collect() and gc.freeze(), so a request
  no longer pays for a collection over objects every request shares. Objects
  from the image are never collected and gc.get_objects() does not list
  them. See docs/python.md.