feat: approximate NVRAM requirement from the CAP analysis (v3.5.15)

GlobalPlatform models this explicitly (GP Card Spec v2.3.1 Table 11-48,
load parameters): C6 = non-volatile code, C7 = volatile data, C8 =
non-volatile data, and 11.5.2.3.7 - when the card makes no code/data
distinction the required minimum is C6 + C8.  The analysis now reports it:

- capmem.memory_json() accepts the load file size (all CAP components -
  the package image the card stores) and returns code.load_file,
  nvram.requirement = load_file + persistent data, and sets the suggested
  C6 to the load file (the bytecode-only Method.cap figure stays in
  code.method_component); _cap_info_body passes the size it already
  computed.
- The CAP estimate box (both the RAM installer and the SCP81
  Install-from-.cap form) leads with "NVRAM requirement ≈ code image +
  data" plus the RAM estimate, keeps the bytecode/other-component split and
  the full data breakdown in Details, and carries the GP citation with the
  caveats (card memory management, allocation rounding and the registry
  entry are not included; the static field image appears in both parts).
- An older server response without code.load_file still renders (the
  bytecode size is used as the fallback).

Tests: Python +1 (load-file semantics) with extended cap-info assertions,
frontend +1 (renderer + fallback).  Help EN/RU, docs/api.md, AGENTS.

587 frontend / 472 Python green; version 3.5.15; sw simple-v268.

frontend/sw.js

@@ -1,4 +1,4 @@
-const CACHE = 'simple-v267';
+const CACHE = 'simple-v268';
 const URLS = [
   'index.html',
   'help.html',