Remove fmt: convert I* files to std::format (part 4)

[brianlball]

Purpose

Continues the remove-fmt campaign (#11553, #11588, #11592). Converts the I*
source files from EnergyPlus::format/fmt::format to C++20 std::format and
replaces the custom {:.NR} flag with standard specifiers.

Scope

File	calls converted
InternalHeatGains.cc	148
IceThermalStorage.cc	7
ICEngineElectricGenerator.cc	4

0 {R}/{T} and 0 EnergyPlus::format remain in these files. The DoubleWrapper /
Fortran-format machinery in IOFiles.hh is left untouched (later PR).

Specifier mapping

Per-field, following the A–H precedent and chosen to preserve the reported precision:

• Magnitude-spanning density columns (e.g. People/Floor Area, the per-area
  Zone Internal Gains Nominal fields) → {:#G}. The old {:.NR} switched to
  scientific below 0.1, so a plain {:.Nf} would truncate small values
  (3.531E-002 → 0.035, losing two sig figs). {:#G} preserves precision
  (0.0353101), matching how A–H handles these EIO "Nominal" tables.
• Always-tiny CO2 rate → {:.4E} — reproduces the baseline mantissa exactly
  (3.8200E-008 → 3.8200E-08, only exponent width differs).
• People counts → {:.1f} — matches the old 1-decimal echo.
• Fractions / DesignLevel watts (always ≥ 0.1) → {:.3f} (exact match).

Verification (energyplus-regressions vs develop, local Windows)

• 0 numeric result diffs — ESO/MTR/SSZ/ZSZ/JSON byte-identical.
• Display-only table/EIO diffs: 543 small + 10 big, 652 EIO.
• The 10 big are irreducible x.x5 people-count flips — old {R} rounds half-up,
  std::format rounds half-to-even. Same benign class accepted in Remove fmt - Part 2: Remove {R}, Files A-D #11588/Remove fmt - Part 3: Remove {R}, Files E-H #11592.
• ⚠️ {:#G} has documented MSVC-vs-gcc trailing-zero divergence; local run is
  Windows-only, so CI (ubuntu/macos) is the authoritative check on the density
  outputs. A–H already ship {:#G} broadly, so risk is low but non-zero.

Indentation note

Whitespace-only line changes are clang-format v19 reflow: std::format is 7 chars
shorter than EnergyPlus::format, so wrapped calls collapse under the 150-col
limit. pre-commit run clang-format passes clean.

Question for review

Is the per-field mapping ({:#G} densities / {:.4E} CO2 / {:.1f} counts) the
approach you want, and are the 10 half-even count-flip diffs acceptable (consistent
with prior parts)?

[github-actions]

⚠️ Regressions detected on ubuntu-24.04 for commit 98e953c

Regression Summary

• EIO: 693
• Table Big Diffs: 10
• Table Small Diffs: 580

• View Results
• Download Regressions

[github-actions]

⚠️ Regressions detected on macos-14 for commit 98e953c

Regression Summary

• EIO: 689
• Table Big Diffs: 10
• Table Small Diffs: 578

• View Results
• Download Regressions

[brianlball]

CI looks good ✅

All checks pass. I went through the regression artifacts from both Linux and macOS.

The 10 "Big Diffs" are all people-count display changes in the
People Internal Gains Nominal table:

Test	change
5ZoneAirCooledWithSpaces / …DaylightingIntMass / …HVAC	11.9 → 11.8
1ZoneEvapCooler_ReportPeriodPowerOutage (+ _naturalVent)	1.0 → 0.9
ASHRAE901_HotelLarge_STD2019_Denver	0.5 → 0.4
ASHRAE901_HotelSmall_STD2019_Denver	0.5 → 0.4, 1.4 → 1.3
TranspiredCollectors	14.3 → 14.2
UserDefinedRoomAirPatterns	1.3 → 1.2
ASHRAE901_OutPatientHealthCare_STD2019_Denver	tiny min/max counts (0.011, 0.056, …) → 0.0 / 0.1

Most are a count landing on a .5 and rounding the other way — old code rounds
half-up, std::format rounds half-to-even. OutPatient is the exception: sub-0.1
per-schedule counts that {:.1f} rounds to 0.0/0.1 where develop printed
scientific notation. All are 1-decimal count display; same class as #11588 / #11592.
(Counts stay {:.1f} deliberately — unlike the densities, {:#G} would render
whole people counts as 5.00000; the sub-0.1 roundings are a negligible
fractional-occupancy artifact.)

The one risk worth checking, checked. These files switch the density columns
to {:#G}, which has a history of printing differently on MSVC vs gcc. I compared
our output across all 689 cases that ran on both Linux and macOS — identical on
every one. No platform divergence.

Why the diff count is on the higher side: densities use {:#G} on purpose to
keep precision on small values — plain fixed formatting would turn 3.531E-002
into 0.035 and lose digits. That trades a larger diff count for no precision loss.

	Linux	macOS
Result diffs	0	0
Table Big	10	10
Table Small	580	578
EIO (echo)	693	689

[github-actions]

⚠️ Regressions detected on macos-14 for commit a19e69d

Regression Summary

• EIO: 690
• Table Big Diffs: 78
• Table Small Diffs: 511

• View Results
• Download Regressions

[github-actions]

⚠️ Regressions detected on ubuntu-24.04 for commit a19e69d

Regression Summary

• EIO: 694
• Table Big Diffs: 78
• Table Small Diffs: 513

• View Results
• Download Regressions